ICAART 2023 Abstracts

Area 1 - Artificial Intelligence
Full Papers Short Papers
Area 2 - Agents
Full Papers Short Papers

Area 1 - Artificial Intelligence

Full Papers
Paper Nr: 7
Title:

Deep W-Networks: Solving Multi-Objective Optimisation Problems with Deep Reinforcement Learning
Authors:

Jernej Hribar, Luke Hackett and Ivana Dusparic
Abstract: In this paper, we build on advances introduced by the Deep Q-Networks (DQN) approach to extend the multiobjective tabular Reinforcement Learning (RL) algorithm W-learning to large state spaces. W-learning algorithm can naturally solve the competition between multiple single policies in multi-objective environments. However, the tabular version does not scale well to environments with large state spaces. To address this issue, we replace underlying Q-tables with DQN, and propose an addition of W-Networks, as a replacement for tabular weights (W) representations. We evaluate the resulting Deep W-Networks (DWN) approach in two widely-accepted multi-objective RL benchmarks: deep sea treasure and multi-objective mountain car. We show that DWN solves the competition between multiple policies while outperforming the baseline in the form of a DQN solution. Additionally, we demonstrate that the proposed algorithm can find the Pareto front in both tested environments.
Download
Paper Nr: 11
Title:

First Go, then Post-Explore: The Benefits of Post-Exploration in Intrinsic Motivation
Authors:

Zhao Yang, Thomas M. Moerland, Mike Preuss and Aske Plaat
Abstract: Go-Explore achieved breakthrough performance on challenging reinforcement learning (RL) tasks with sparse rewards. The key insight of Go-Explore was that successful exploration requires an agent to first return to an interesting state (‘Go’), and only then explore into unknown terrain (‘Explore’). We refer to such exploration after a goal is reached as ‘post-exploration’. In this paper, we present a clear ablation study of post-exploration in a general intrinsically motivated goal exploration process (IMGEP) framework, that the Go-Explore paper did not show. We study the isolated potential of post-exploration, by turning it on and off within the same algorithm under both tabular and deep RL settings on both discrete navigation and continuous control tasks. Experiments on a range of MiniGrid and Mujoco environments show that post-exploration indeed helps IMGEP agents reach more diverse states and boosts their performance. In short, our work suggests that RL researchers should consider using post-exploration in IMGEP when possible since it is effective, method-agnostic, and easy to implement.
Download
Paper Nr: 26
Title:

Reinforcement Learning Explained via Reinforcement Learning: Towards Explainable Policies through Predictive Explanation
Authors:

Léo Saulières, Martin C. Cooper and Florence Bannay
Abstract: In the context of reinforcement learning (RL), in order to increase trust in or understand the failings of an agent’s policy, we propose predictive explanations in the form of three scenarios: best-case, worst-case and most-probable. After showing W[1]-hardness of finding such scenarios, we propose linear-time approximations. In particular, to find an approximate worst/best-case scenario, we use RL to obtain policies of the environment viewed as a hostile/favorable agent. Experiments validate the accuracy of this approach.
Download
Paper Nr: 30
Title:

Adding Time and Subject Line Features to the Donor Journey
Authors:

Greg Lee, Ajith K. Raghavan and Mark Hobbs
Abstract: The donor journey is the path a charitable constituent takes on their way to making a donation. Charities are moving towards more electronic communication and most appeals are now sent via email. The donor journey can be followed electronically, monitoring constituent and charity actions. Previous research has shown that it is possible to use past actions of a donor to predict their next gift within $25. We build on this research by adding new features that capture the time between actions, as well as new email features, including subject lines features in such a way as to isolate their effect on model accuracy. These additions show a small improvement in accuracy of recurrent neural network models for most charities, showing these features do indeed help deep learning methods understand the donor journey.
Download
Paper Nr: 31
Title:

Identifying Visitor's Paintings Appreciation for AI Audio Guide in Museums
Authors:

Mari Saito, Takato Okudo, Makoto Yamada and Seiji Yamada
Abstract: This paper describes an application of machine learning for predicting whether a user is engaged in art appreciation to develop AI audio guide systems that can automatically control when guidance is provided. Although many studies on intelligent audio guides in museums have been done, there are few that have tried to develop AI audio guide systems that can begin to play audio guides automatically when visitors are engaged in art appreciation. In this paper, we determine the timing at which to begin an audio guide by classifying two classes, that is, whether the user is engaged in art appreciation or not, which is identified at the museum. We apply supervised machine learning for time-series data to the classification. We conducted experiments with participants in a real museum and collected labeled time-series data of participants heads’ postures and movements as training data. Then, we applied a classification learning algorithm for time-series data to predict when participants were involved in painting appreciation, executed model selection, and experimentally evaluated the models with the collected data. Since the results showed a good accuracy of over 82%, we confirmed that our machine learning-based approach to real-time identification of painting appreciation is promising for AI audio guide systems.
Download
Paper Nr: 42
Title:

Deep Learning Model Selection With Parametric Complexity Control
Authors:

Olga Grebenkova, Oleg Bakhteev and Vadim Strijov
Abstract: The paper is devoted to deep learning model complexity. It is estimated by Bayesian inference and based on a computational budget. The idea of the proposed method is to represent deep learning model parameters in the form of hypernetwork output. A hypernetwork is a supplementary model which generates parameters of the selected model. This paper considers the minimum description length from a Bayesian point of view. We introduce prior distributions of deep learning model parameters to control the model complexity. The paper analyzes and compares three types of regularization to define the parameter distribution. It infers and generalizes the model evidence as a criterion that depends on the required model complexity. Finally, it analyzes this method in the computational experiments on the Wine, MNIST, and CIFAR-10 datasets.
Download
Paper Nr: 43
Title:

Innovations and Emerging Technologies: A Study of the Italian Intellectual Property Knowledge Database
Authors:

Annamaria Demarinis Loiotile, Francesco De Nicolò, Alfonso Monaco, Sabina Tangaro, Shiva Loccisano, Giuseppe Conti, Adriana Agrimi, Nicola Amoroso and Roberto Bellotti
Abstract: A great mine of innovation is represented by the excellence of the scientific know-how of the Italian universities and research centers. But very often university patents remain unvalued and unexploited, in the so-called “Valley of death”. In the framework of Intellectual Property Analytics and Patent Informatics, this paper analyses the Italian patent database “Knowledge Share” and its proposed classifications (10 technological areas). By means of Natural Language Processing (NLP) techniques, we examined 1694 patents from 89 Italian Research Institutions and a cluster analysis revealed the existence of 8 homogeneous clusters instead of the 10 proposed by the platform. Thus, our findings suggest the presence of possible inhomogeneities within the traditional classifications, probably due to the emergence of novel technologies or cross-domain areas, e.g., Healthcare 4.0; moreover, these clusters could lead to better performance in terms of offer/demand matching for the platform users.
Download
Paper Nr: 46
Title:

Towards a Foundation for Intelligent Contracts
Authors:

Georgios Stathis, Athanasios Trantas, Giulia Biagioni, Jaap van den Herik, Bart Custers, Laura Daniele and Theofilos Katsigiannis
Abstract: This article investigates the incorporation of Artificial Intelligence (AI) within LegalTech. We define an ontology to form the basis for Trustworthy AI processing of contract automation. The value of our research is that it applies an ontology, as existing tool, in contract automation. Two perspectives are emphasized: communications analysis and risk analysis. They are explored under a new prism. Our context is Intelligent Contracts (iContracts), which aim at reducing the time-consuming and often complex contractual process by minimizing human involvement. Contract communications and risk analysis processes are often neglected in automation. Therefore, our research investigates to what extent is possible to design an ontology for contract automation based upon the combination of both. Our methodology is twofolded. First, we concentrate on applying key word search on an online database to demonstrate the lack of available solutions. Second, we develop an ontology based upon a case study of a freelancer agreement. Of course, we use existing literature to further engineer the ontology. Our finding shows that 9.4 percent of LegalTech solutions deal with contract automation. From them, 0.7 percent focus on communications and risk automation for contracting. The conceptual expressiveness of the ontology is validated with research to the use case. A follow-up discussion suggests that the ontology should be further engineered from a third perspective, trustworthiness, and should be re-validated experimentally. Our conclusion underlines the need for further innovation in contract automation, especially in relation to communications and risk data.
Download
Paper Nr: 50
Title:

Generation of Facial Images Reflecting Speaker Attributes and Emotions Based on Voice Input
Authors:

Kotaro Koseki, Yuichi Sei, Yasuyuki Tahara and Akihiko Ohsuga
Abstract: The task of “face generation from voice” will bring about a significant change in the way voice calls are made. Voice calls create a psychological gap compared to face to face communication because the other party’s face is not visible. Generating a face from voice can alleviate this psychological gap and contribute to more efficient communication. Multimodal learning is a machine learning method that uses different data (e.g., voice and face images) and is being studied to combine various types of information such as text, images, and voice, as in google’s imagen(Saharia et al., 2022). In this study, we perform multimodal learning of speech and face images using a CNN convolutional speech encoder and a face image variational autoencoder (VAE: Variational Autoencoder) to create models that can represent speech and face images of different modalities in the same latent space. Focusing on the emotional information of speech, we also built a model that can generate face images that reflect the speaker’s emotions and attributes in response to input speech. As a result, we were able to generate face images that reflect rough emotions and attributes, although there are variations in the emotions depending on the type of emotion.
Download
Paper Nr: 53
Title:

Coverage-Guided Fuzzing for Plan-Based Robotics
Authors:

Tim Meywerk, Vladimir Herdt and Rolf Drechsler
Abstract: Autonomous robots are used increasingly in dynamic and safety-critical environments. In these environments the correctness of the robotic plan is of utmost importance. In many other domains, coverage-guided fuzzing has proven to be an effective way to ensure the correctness of software programs. In coverage-guided fuzzing, inputs to a program are generated semi-randomly and the correctness of the output is checked automatically. This way, a large number of test cases can be run without manual interaction. In this work we present our approach to coverage-guided fuzzing for plan-based robotics and our prototypical implementation for the planning language CPL. We also introduce a novel coverage metric for the domain of plan-based robotics.
Download
Paper Nr: 58
Title:

Data Augmentation Through Expert-Guided Symmetry Detection to Improve Performance in Offline Reinforcement Learning
Authors:

Giorgio Angelotti, Nicolas Drougard and Caroline C. Chanel
Abstract: Offline estimation of the dynamical model of a Markov Decision Process (MDP) is a non-trivial task that greatly depends on the data available in the learning phase. Sometimes the dynamics of the model is invariant with respect to some transformations of the current state and action. Recent works showed that an expert-guided pipeline relying on Density Estimation methods as Deep Neural Network based Normalizing Flows effectively detects this structure in deterministic environments, both categorical and continuous-valued. The acquired knowledge can be exploited to augment the original data set, leading eventually to a reduction in the distributional shift between the true and the learned model. Such data augmentation technique can be exploited as a preliminary process to be executed before adopting an Offline Reinforcement Learning architecture, increasing its performance. In this work we extend the paradigm to also tackle non-deterministic MDPs, in particular, 1) we propose a detection threshold in categorical environments based on statistical distances, and 2) we show that the former results lead to a performance improvement when solving the learned MDP and then applying the optimized policy in the real environment.
Download
Paper Nr: 67
Title:

A New Dynamic Community-Based Recommender System
Authors:

Sabrine Ben Abdrabbah, Nahla Ben Amor and Raouia Ayachi
Abstract: Due to the rapid changes in users’ preferences over time, it becomes increasingly important to focus on the temporal evolution of the users’ behavioral patterns to capture the most relevant items to the users. This paper proposes a new framework for dynamic and overlapping community-based collaborative filtering, which models at first the dynamic behavior of users’ interests into a temporal network of items. Then, we take advantage of the dynamic and overlapping community detection techniques to find the best partition of similar items. The advantage of doing so is to (i) avoid processing the entire system to select similar items and eventually overcome the scalability problem and (ii) provide users with a recommendation of items similar to the latest appreciated items to match the current users’ taste and avoid consequently sparse cases. We conduct experiments to study the sensitivity of some parameters (e.g., the datasets and the similarity measures) on the recommendations’ quality. Experimental results show a considerable improvement in the proposed framework recommendation’s accuracy compared to the state-of-the-art recommender systems.
Download
Paper Nr: 82
Title:

Trade-off Clustering Approach for Multivariate Multi-Step Ahead Time-Series Forecasting
Authors:

Konstandinos Aiwansedo, Wafa Badreddine and Jérôme Bosche
Abstract: Time-Series forecasting has gained a lot of steam in recent years. With the advent of Big Data, a considerable amount of data is more available across multiple fields, thus providing an opportunity for processing historical business-oriented data in an attempt to predict trends, identify changes and inform strategic decision-making. The abundance of time-series data has prompted the development of state-of-the-art machine learning algorithms, such as neural networks, capable of forecasting both univariate and multivariate time-series data. Various time-series forecasting approaches can be implemented when leveraging the potential of deep neural networks. Determining the upsides and downsides of each approach when presented with univariate or multivariate time-series data, thus becomes a crucial matter. This evaluation focuses on three forecasting approaches: a single model forecasting approach (SMFA), a global model forecasting model (GMFA) and a cluster-based forecasting approach (CBFA). The study highlights the fact that the decision pertaining to the finest forecasting approach often is a question of trade-off between accuracy, execution time and dataset size. In this study, we also compare the performance of 6 deep learning architectures when dealing with both univariate and multivariate time-series datasets for multi-step ahead time-series forecasting, across 6 benchmark datasets.
Download
Paper Nr: 86
Title:

Rewiring Knowledge Graphs by Graph Neural Network Link Predictions
Authors:

Alex Romanova
Abstract: Knowledge Graphs recently received increasing attention from academia and industry as a new era in data-driven technology. By building relationships graphs are ’connecting the dots’ and moving data from zerodimensional to multi-dimensional space. Emerging Graph Neural Network (GNN) models are building a bridge between graph topology and deep learning. In this study we examine how to use GNN link prediction models to rewire knowledge graphs and detect unexplored relationships between graph nodes. We investigate diverse advantages of using highly connected and highly disconnected node pairs for graph mining techniques.
Download
Paper Nr: 88
Title:

Interactive Video Saliency Prediction: The Stacked-convLSTM Approach
Authors:

N. Wondimu, U. Visser and C. Buche
Abstract: Cognitive and neuroscience of attention researches suggest the use of spatio-temporal features for an efficient video saliency prediction. This is due to the representative nature of spatio-temporal features for data collected across space and time, such as videos. Video saliency prediction aims to find visually salient regions in a stream of images. Many video saliency prediction models are proposed in the past couple of years. Due to the unique nature of videos from that of static images, the earliest efforts to employ static image saliency prediction models for video saliency prediction task yield reduced performance. Consequently, dynamic video saliency prediction models that use spatio-temporal features were introduced. These models, especially deep learning based video saliency prediction models, transformed the state-of-the-art of video saliency prediction to a better level. However, video saliency prediction still remains a considerable challenge. This has been mainly due to the complex nature of video saliency prediction and scarcity of representative saliency benchmarks. Given the importance of saliency identification for various computer vision tasks, revising and enhancing the performance of video saliency prediction models is crucial. To this end, we propose a novel interactive video saliency prediction model that employs stacked-ConvLSTM based architecture along with a novel XY-shift frame differencing custom layer. Specifically, we introduce an encoder-decoder based architecture with a prior layer undertaking XY-shift frame differencing, a residual layer fusing spatially processed (VGG-16 based) features with XY-shift frame differenced frames, and a stacked-ConvLSTM component. Extensive experimental results over the largest video saliency dataset, DHF1K, show the competitive performance of our model against the state-of-the-art models.
Download
Paper Nr: 89
Title:

LSTM, ConvLSTM, MDN-RNN and GridLSTM Memory-based Deep Reinforcement Learning
Authors:

Fernando F. Duarte, Nuno Lau, Artur Pereira and Luís P. Reis
Abstract: Memory-based Deep Reinforcement Learning has been shown to be a viable solution to successfully learn control policies directly from high-dimensional sensory data in complex vision-based control tasks. At the core of this success lies the Long Short-Term Memory or LSTM, a well-known type of Recurrent Neural Network. More recent developments have introduced the ConvLSTM, a convolutional variant of the LSTM and the MDN-RNN, a Mixture Density Network combined with an LSTM, as memory modules in the context of Deep Reinforcement Learning. The defining characteristic of the ConvLSTM is its ability to preserve spatial information, which may prove to be a crucial factor when dealing with vision-based control tasks while the MDN-RNN can act as a predictive memory eschewing the need to explicitly plan ahead. Also of interest to this work is the GridLSTM, a network of LSTM cells arranged in a multidimensional grid. The objective of this paper is therefore to perform a comparative study of several memory modules, based on the LSTM, ConvLSTM, MDN-RNN and GridLSTM in the scope of Deep Reinforcement Learning, and more specifically as the memory modules of the agent. All experiments were validated using the Atari 2600 videogame benchmark.
Download
Paper Nr: 94
Title:

Predicting Off-Block Delays: A Case Study at Paris - Charles de Gaulle International Airport
Authors:

Thibault Falque, Bertrand Mazure and Karim Tabia
Abstract: Punctuality is a sensitive issue in large airports and hubs for passenger experience and for controlling operational costs. This paper presents a real and challenging problem of predicting and explaining flight off-block delays. We study the case of the international airport Paris Charles de Gaulle (Paris-CDG) starting from the specificities of this problem at Paris-CDG until the proposal of modelings then solutions and the analysis of the results on real data covering an entire year of activity. The proof of concept provided in this paper allows us to believe that the proposed approach could help improving the management of delays and reduce the impact of the resulting consequences.
Download
Paper Nr: 98
Title:

Using NLP to Enrich Scientific Knowledge Graphs: A Case Study to Find Similar Papers
Authors:

Xavier Quevedo and Janneth Chicaiza
Abstract: In recent years, Knowledge Graphs have become increasingly popular thanks to the potential of Semantic Web technologies and the development of NoSQL graph-based. A knowledge graph that describes scholarly production makes the literature metadata legible for machines. Making the paper’s text legible for machines enables them to discover and leverage relevant information for the scientific community beyond searching based on metadata fields. Thus, scientific knowledge graphs can become catalysts to drive research. In this research, we reuse an existing scientific knowledge graph and enrich it with new facts to demonstrate how this information can be used to improve tasks like finding similar documents. To identify new entities and relationships we combine two different approaches: (1) an RDF scheme-based approach to recognize named entities, and (2) a sequence labeler based on spaCy to recognize entities and relationships on papers’ abstracts. Then, we compute the semantic similarity among papers considering the original graph and the enriched one to state what is the graph that returns the closest similarity. Finally, we conduct an experiment to verify the value or contribution of the additional information, i.e. new triples, obtained by analyzing the content of the abstracts of the papers.
Download
Paper Nr: 103
Title:

Quantum Reinforcement Learning for Solving a Stochastic Frozen Lake Environment and the Impact of Quantum Architecture Choices
Authors:

Theodora-Augustina Drăgan, Maureen Monnet, Christian B. Mendl and Jeanette M. Lorenz
Abstract: Quantum reinforcement learning (QRL) models augment classical reinforcement learning schemes with quantum-enhanced kernels. Different proposals on how to construct such models empirically show a promising performance. In particular, these models might offer a reduced parameter count and shorter times to reach a solution than classical models. It is however presently unclear how these quantum-enhanced kernels as subroutines within a reinforcement learning pipeline need to be constructed to indeed result in an improved performance in comparison to classical models. In this work we exactly address this question. First, we propose a hybrid quantum-classical reinforcement learning model that solves a slippery stochastic frozen lake, an environment considerably more difficult than the deterministic frozen lake. Secondly, different quantum architectures are studied as options for this hybrid quantum-classical reinforcement learning model, all of them well-motivated by the literature. They all show very promising performances with respect to similar classical variants. We further characterize these choices by metrics that are relevant to benchmark the power of quantum circuits, such as the entanglement capability, the expressibility, and the information density of the circuits. However, we find that these typical metrics do not directly predict the performance of a QRL model.
Download
Paper Nr: 117
Title:

Qualitative Spatial Representation and Reasoning About Fold Strata
Authors:

Yuta Taniuchi and Kazuko Takahashi
Abstract: We propose a method of handling strata in qualitative spatial representation. We make a model for a typical fold structure projected onto a two-dimensional plane extracted by a rectangle. It is expressed by a pair of sequences of symbols that represents the strata configuration and the shape of the layers, respectively. We define the validity required of the representation and show that the representation and the model have a one-to-one relation. Moreover, we define operations on the representation, such as rotation and symmetric transitions, and show that validity is preserved. We also show that global data can be constructed by connecting local data. This method can provide a logical explanation of the processes involved in strata-generation prediction, which in the field of structural geology have been examined manually to date, and find results that manual analysis may overlook.
Download
Paper Nr: 119
Title:

Automatic Facility Layout Design System Using Deep Reinforcement Learning
Authors:

Hikaru Ikeda, Hiroyuki Nakagawa and Tatsuhiro Tsuchiya
Abstract: Facility layout designing aims to deploy functional objects in appropriate locations within the logistics facilities and production facilities. The designer’s ability to create a layout is a major factor in the quality of the layout because they need to satisfy functional requirements like lead time, relations among functional objects to deploy and material handling costs. In this paper, a deep reinforcement learning (RL) based automatic layout design system is developed. Deep Q-Networt (DQN) is introduced to solve facility layout problem (FLP) by the adaptability of RL with the expression of deep neural networks. We apply the developed system to the existing FLP and compare the layout result with conventional RL based system. Consequently, the performance improvement was confirmed in terms of the relations among units in the created layout comparing to the RL based system.
Download
Paper Nr: 138
Title:

Partner Selection Strategy in Open, Dynamic and Sociable Environments
Authors:

Qin Liang, Wen Gu, Shohei Kato, Fenghui Ren, Guoxin Su, Takayuki Ito and Minjie Zhang
Abstract: In multi-agent systems, agents with limited capabilities need to find a cooperation partner to accomplish complex tasks. Evaluating the trustworthiness of potential partners is vital in partner selection. Current approaches are mainly averaged-based, aggregating advisors’ information on partners. These methods have limitations, such as vulnerability to unfair rating attacks, and may be locally convergent that cannot always select the best partner. Therefore, we propose a ranking-based partner selection (RPS) mechanism, which clusters advisors into groups according to their ranking of trustees and gives recommendations based on groups. Besides, RPS is an online-learning method that can adjust model parameters based on feedback and evaluate the stability of advisors’ ranking behaviours. Experiments demonstrate that RPS performs better than state-of-the-art models in dealing with unfair rating attacks, especially when dishonest advisors are the majority.
Download
Paper Nr: 151
Title:

Multi-Label Learning for Aspect Category Detection of Arabic Hotel Reviews Using AraBERT
Authors:

Asma Ameur, Sana Hamdi and Sadok Ben Yahia
Abstract: Studying people’s satisfaction with social media is vital to understanding the users’ needs. Nowadays, textual hotel reviews are used to evaluate the hotel’s e-reputation. In this context, we are interested in Aspect Category Detection (ACD) as a subtask of aspect-based sentiment analysis. This task needs to be investigated through multi-label classification, which is more challenging, in natural language processing, than single-label classification. Our study leverages the potential of transfer learning with the pre-trained AraBERT model for contextual text representation. We are based on the Arabic SemEval-2016 data set for hotel reviews. We propose a specific preprocessing for this Arabic reviews dataset to improve the performance. In addition, as this data suffers from an imbalanced distribution, we use a dynamically weighted loss function approach to deal with imbalanced classes. The carried-out results outperform the pioneering state-of-the-art of the Arabic ACD with an F1 score of 67:3%.
Download
Paper Nr: 153
Title:

Improving Convergence for Quantum Variational Classifiers Using Weight Re-Mapping
Authors:

Michael Kölle, Alessandro Giovagnoli, Jonas Stein, Maximilian B. Mansky, Julian Hager and Claudia Linnhoff-Popien
Abstract: In recent years, quantum machine learning has seen a substantial increase in the use of variational quantum circuits (VQCs). VQCs are inspired by artificial neural networks, which achieve extraordinary performance in a wide range of AI tasks as massively parameterized function approximators. VQCs have already demonstrated promising results, for example, in generalization and the requirement for fewer parameters to train, by utilizing the more robust algorithmic toolbox available in quantum computing. A VQCs’ trainable parameters or weights are usually used as angles in rotational gates and current gradient-based training methods do not account for that. We introduce weight re-mapping for VQCs, to unambiguously map the weights to an interval of length 2π, drawing inspiration from traditional ML, where data rescaling, or normalization techniques have demonstrated tremendous benefits in many circumstances. We employ a set of five functions and evaluate them on the Iris and Wine datasets using variational classifiers as an example. Our experiments show that weight re-mapping can improve convergence in all tested settings. Additionally, we were able to demonstrate that weight re-mapping increased test accuracy for the Wine dataset by 10% over using unmodified weights.
Download
Paper Nr: 159
Title:

Exotic Bets: Evolutionary Computing Coupled with Bet Mechanisms for Model Selection
Authors:

Simon Reichhuber and Sven Tomforde
Abstract: This paper presents a framework for the application of an external bet-based evolutionary algorithm to the problem of model selection. In particular, we have defined two new risk functions, called sample space exoticness and configuration space exoticness. The latter is used to manage the risk of bet placement. Further, we explain how to implement the bet-based approach for model selection in the domain of multi-class classification and experimentally compare the performance of the algorithm to reference derivative-free hyperparameter optimisers (GA and Bayesian Optimisation) on MNIST. Finally, we experimentally show that for the classifiers SVM, MLP, and Nearest Neighbors the balanced accuracy can be increased by up to three percentage points.
Download
Paper Nr: 162
Title:

An Application of Priority-Based Lightweight Ontology Merging
Authors:

Rim Mohamed, Truong-Thanh Ma and Zied Bouraoui
Abstract: Merging multiple and frequently contradictory sources of information has been identified as a significant issue in the semantic web community. In addition, pieces of information to be combined are provided with uncertainty due, for instance, to the reliability of sources. To solve this, possibility theory offers a useful tool for representing and reasoning with uncertain, partial, and inconsistent information. In this paper, we concentrate on dance video processing, in which many inconsistent information sources exist. Therefore, we propose possibilistic merging operators for the dance OWL2-EL ontologies to deal with the conflicting dance sources. We represent an extension of EL within a possibility theory setting. It leverages a min-based operator to merge the ontologies based on possible distributions. Furthermore, the semantic fusion of these distributions has a natural syntactic counterpart when dealing with EL ontologies. The min-based fusion operator is recommended when distinct dance sources that provide information are dependent.
Download
Paper Nr: 166
Title:

A Fine-Tuning Aggregation Convolutional Neural Network Surrogate Model of Strategy Selecting Mechanism for Repeated-Encounter Bilateral Automated Negotiation
Authors:

Shengbo Chang and Katsuhide Fujita
Abstract: Negotiation with the same opponent for multiple times for each in a different domain commonly occurs in real life. We consider this automated negotiation problem as repeated-encounter bilateral automated negotiation (RBAN), in which it is essential to learn experiences from the history of coping with the opponent. This study presents a surrogate-model-based strategy selecting mechanism that learns experiences in RBAN by fine-tuning the proposed aggregation convolutional neural network (CNN) surrogate model (ACSM). ACSM is promised to assess strategies more precisely by applying CNN to extract features from a matrix showing the outcomes’ utility distribution. It ensures the abundance of extracted features by aggregating multiple CNNs trained with diverse opponents. The fine-tuning approach adapts ACSM to the opponent in RBAN by feeding the present negotiation results to ACSM. We evaluate ACSM and the fine-tuning approach experimentally by selecting a strategy for a time-dependent agent. The experiments of negotiating with four Automated Negotiating Agents Competition (ANAC) champions and six basic agents are performed. ACSM is tested on 600 negotiation scenarios originating from ANAC domains. The fine-tuning approach is tested on 60 RBNA sessions. The experimental results indicate that ACSM outperforms an existing feature-based surrogate model, and the fine-tuning approach is able to adapt ACSM to the opponent in RBAN.
Download
Paper Nr: 172
Title:

Deep Hybrid Bagging Ensembles for Classifying Histopathological Breast Cancer Images
Authors:

Fatima-Zahrae Nakach, Ali Idri and Hasnae Zerouaoui
Abstract: This paper proposes the use of transfer learning and ensemble learning for binary classification of breast cancer histological images over the four magnification factors of the BreakHis dataset: 40×, 100×, 200× and 400×. The proposed bagging ensembles are implemented using a set of hybrid architectures that combine pre-trained deep learning techniques for feature extraction with machine learning classifiers as base learners (MLP, SVM and KNN). The study evaluated and compared: (1) bagging ensembles with their base learners, (2) bagging ensembles with a different number of base learners (3, 5, 7 and 9), (3) single classifiers with the best bagging ensembles, and (4) best bagging ensembles of each feature extractor and magnification factor. The best cluster of the outperforming models was chosen using the Scott Knott (SK) statistical test, and the top models were ranked using the Borda Count voting system. The best bagging ensemble achieved a mean accuracy value of 93.98%, and was constructed using 3 base learners, 200× as a magnification factor, MLP as a classifier, and DenseNet201 as a feature extractor. The results demonstrated that bagging hybrid deep learning is an effective and a promising approach for the automatic classification of histopathological breast cancer images.
Download
Paper Nr: 181
Title:

EMTE: An Enhanced Medical Terms Extractor Using Pattern Matching Rules
Authors:

Monah B. Hatoum, Jean-Claude Charr, Christophe Guyeux, David Laiymani and Alia Ghaddar
Abstract: Downstream tasks like clinical textual data classification perform best when given good-quality datasets. Most of the existing clinical textual data preparation techniques rely on two main approaches, removing irrelevant data using cleansing techniques or extracting valuable data using feature extraction techniques. However, they still have limitations, mainly when applied to real-world datasets. This paper proposes a cleansing approach (called EMTE) which extracts phrases (medical terms, abbreviations, and negations) using pattern-matching rules based on the linguistic processing of the clinical textual data. Without requiring training, EMTE extracts valuable medical data from clinical textual records even if they have different writing styles. Furthermore, since EMTE relies on dictionaries to store abbreviations and pattern-matching rules to detect phrases, it can be easily maintained and extended for industrial use. To evaluate the performance of our approach, we compared the performance of EMTE to three other techniques. All four cleansing techniques were applied to a large industrial imbalanced dataset, consisting of 2.21M samples from different specialties with 1,050 ICD-10 codes. The experimental results on several Deep Neural Network (DNN) algorithms showed that our cleansing approach significantly improves the trained models’ performance compared to the other tested techniques and according to different metrics.
Download
Paper Nr: 199
Title:

An Audit Framework for Technical Assessment of Binary Classifiers
Authors:

Debarati Bhaumik and Diptish Dey
Abstract: Multilevel models using logistic regression (MLogRM) and random forest models (RFM) are increasingly deployed in industry for the purpose of binary classification. The European Commission’s proposed Artificial Intelligence Act (AIA) necessitates, under certain conditions, that application of such models is fair, transparent, and ethical, which consequently implies technical assessment of these models. This paper proposes and demonstrates an audit framework for technical assessment of RFMs and MLogRMs by focussing on model-, discrimination-, and transparency & explainability-related aspects. To measure these aspects 20 KPIs are proposed, which are paired to a traffic light risk assessment method. An open-source dataset is used to train a RFM and a MLogRM model and these KPIs are computed and compared with the traffic lights. The performance of popular explainability methods such as kernel- and tree-SHAP are assessed. The framework is expected to assist regulatory bodies in performing conformity assessments of binary classifiers and also benefits providers and users deploying such AI-systems to comply with the AIA.
Download
Paper Nr: 232
Title:

A Robust Adaptive Workload Orchestration in Pure Edge Computing
Authors:

Zahra Safavifar, Charafeddine Mechalikh and Fatemeh Golpayegani
Abstract: Pure Edge computing (PEC) aims to bring cloud applications and services to the edge of the network to support the growing user demand for time-sensitive applications and data-driven computing. However, mobility and limited computational capacity of edge devices pose challenges in supporting some urgent and computationally intensive tasks with strict response time demands. If the execution results of these tasks exceed the deadline, they become worthless and can cause severe safety issues. Therefore, it is essential to ensure that edge nodes complete as many latency-sensitive tasks as possible. In this paper, we propose a Robust Adaptive Workload Orchestration (R-AdWOrch) model to minimize deadline misses and data loss by using priority definition and a reallocation strategy. The results show that R-AdWOrch can minimize deadline misses of urgent tasks while minimizing the data loss of lower priority tasks under all conditions.
Download
Paper Nr: 245
Title:

InsectDSOT: A Neural Network for Insect Detection in Olive Trees
Authors:

Lotfi Souifi, Afef Mdhaffar, Ismael B. Rodriguez, Mohamed Jmaiel and Bernd Freisleben
Abstract: Controlling insect pests in agricultural fields is a major concern. Despite technological developments, most farm management methods and technologies still rely on experts for management and do not yet match the criteria required for precise insect pest control. In this paper, we present a neural network approach for detecting and counting insects. Using the Yolov5n 6.1 version as a baseline model, this paper proposes replacing the Conv layers in the original model’s backbone and neck with the RepVGG layer. We use transfer learning to improve performance by training our proposal on the MS COCO dataset and then use the output model of this training as the input of our new training. Our proposal is validated using the DIRT (Dacus Image Recognition Toolkit) dataset. The obtained results demonstrate that our approach, based on an improved Yolov5, achieves 86.1% of precision. It outperforms four versions of the original yolov5 and yolov5-based versions with modified backbones based on lightweight models.
Download
Paper Nr: 251
Title:

An Ontology-Based Augmented Observation for Decision-Making in Partially Observable Environments
Authors:

Saeedeh Ghanadbashi, Akram Zarchini and Fatemeh Golpayegani
Abstract: Decision-making is challenging for agents operating in partially observable environments. In such environments, agents’ observation is often based on incomplete, ambiguous, and noisy sensed data, which may lead to perceptual aliasing. This means there might be distinctive states of the environment that appear the same to the agents, and agents fail to take suitable actions. Currently, machine learning, collaboration, and practical reasoning techniques are used to improve agents’ observation and their performance in such environments. However, their long exploration and negotiation periods make them incapable of reacting in real time and making decisions on the fly. The Ontology-based Observation Augmentation Method (OOAM) proposed here, improves agents’ action selection in partially observable environments using domain ontology. OOAM generates an ontology-based schema (i.e., mapping low-level sensor data to high-level concepts), and infers implicit observation data from explicit ones. OOAM is evaluated in a job shop scheduling environment, where the required sensed data to process the orders can be delayed or corrupted. The results show that the average utilization rate and the total processed orders have increased by 17% and 25% respectively compared to Trust Region Policy Optimization (TRPO) as a state-of-the-art method.
Download
Paper Nr: 256
Title:

Studying Narrative Economics by Adding Continuous-Time Opinion Dynamics to an Agent-Based Model of Co-Evolutionary Adaptive Financial Markets
Authors:

Arwa Bokhari and Dave Cliff
Abstract: In 2017 Robert Shiller, a Nobel Laureate, introduced Narrative Economics, an approach to explaining aspects of economies that are difficult to comprehend when analyzed using conventional methods: in light of narratives (i.e., stories) that participants in asset markets hear, believe, and tell each other, some observable economic factors, such as price dynamics of otherwise valueless digital assets, can be explained largely within the context of those narratives. As Shiller argues, it is best to explain and understand seemingly irrational and hard-to-explain behaviors, such as investing in highly volatile cryptocurrency markets, in narrative terms: people invest because they believe that it makes sense to do so, or have a heartfelt opinion about the prospects of the asset, and they share these beliefs and opinions with themselves and others in the form of narratives. In this paper, we address the question of how an agent-based modeling platform can be developed to be used for studying narrative economics. To do this, we integrate two very recently published developments. From the field of agent-based models of financial markets, we use the PRDE adaptive zero-intelligence trader strategy introduced by Cliff (2022), and we extend it to integrate a continuous-time real-valued nonlinear opinion dynamics model reported by Bizyaeva et al. (2022). In our integrated system, each trader holds an opinion variable whose value can be altered by interaction with other agents, modeling the influence that narratives have on an agent’s opinions, and which can also be altered by observation of events in the market. Furthermore, the PRDE algorithm is modified to allow each trader’s trading behavior to smoothly alter as that trader’s opinion dynamically varies. Results reported for the first time here show that in our model there is a tightly coupled circular interplay between opinions and prices: changes in the distribution of opinions can affect subsequent price dynamics; and changes in price dynamics can affect the consequent distribution of opinions. Thus this paper presents a first demonstration of the reliability and effectiveness of our new agent-based modeling platform for use in studying issues in narrative economics. Python source-code for our model is being made freely available as open-source release on GitHub, to allow other researchers to replicate and extend our work.
Download
Paper Nr: 265
Title:

Quantitative Performance Analysis from Discrete Perspective: A Case Study of Chip Detection in Turning Process
Authors:

Sharmin Sultana Sheuly, Mobyen Uddin Ahmed and Shahina Begum
Abstract: Good performance of the Machine Learning (ML) model is an important requirement associated with ML-integrated manufacturing. An increase in performance improvement methods such as hyperparameter tuning, data size increment, feature extraction, and architecture change leads to random attempts while improving performance. This can result in unnecessary consumption of time and performance improvement solely depending on luck. In the proposed study, a quantitative performance analysis on the case study of chip detection is performed from six perspectives: hyperparameter change, feature extraction method, data size increment, and concatenated Artificial Neural Network (ANN) architecture. The focus of the analysis is to create a consolidated knowledge of factors affecting ML model performance in turning process quality prediction. Metal peels such as chips are designed at the time of metal cutting (turning process) and the shape of these chips indicates the quality of the turning process. The result of the proposed study shows that following a fixed recipe does not always improve performance. In the case of performance improvement, data quality plays the main role. Additionally, the choice of an ML algorithm and hyperparameter tuning plays an essential role in performance.
Download
Paper Nr: 266
Title:

Farsighter: Efficient Multi-Step Exploration for Deep Reinforcement Learning
Authors:

Yongshuai Liu and Xin Liu
Abstract: Uncertainty-based exploration in deep reinforcement learning (RL) and deep multi-agent reinforcement learning (MARL) plays a key role in improving sample efficiency and boosting total reward. Uncertainty-based exploration methods often measure the uncertainty (variance) of the value function; However, existing exploration strategies either underestimate the uncertainty by only considering the local uncertainty of the next immediate reward or estimate the uncertainty by propagating the uncertainty for all the remaining steps in an episode. Neither approach can explicitly control the bias-variance trade-off of the value function. In this paper, we propose Farsighter, an explicit multi-step uncertainty exploration framework. Specifically, Farsighter considers the uncertainty of exact k future steps and it can adaptively adjust k. In practice, we learn Bayesian posterior over Q-function in discrete cases and over action in continuous cases to approximate uncertainty in each step and recursively deploy Thompson sampling on the learned posterior distribution with TD(k) update. Our method can work on general tasks with high/low-dimensional states, discrete/continuous actions, and sparse/dense rewards. Empirical evaluations show that Farsighter outperforms SOTA explorations on a wide range of Atari games, robotic manipulation tasks, and general RL tasks.
Download
Paper Nr: 268
Title:

Interpretable Machine Learning for Modelling and Explaining Car Drivers' Behaviour: An Exploratory Analysis on Heterogeneous Data
Authors:

Mir Riyanul Islam, Mobyen Uddin Ahmed and Shahina Begum
Abstract: Understanding individual car drivers’ behavioural variations and heterogeneity is a significant aspect of developing car simulator technologies, which are widely used in transport safety. This also characterizes the heterogeneity in drivers’ behaviour in terms of risk and hurry, using both real-time on-track and in-simulator driving performance features. Machine learning (ML) interpretability has become increasingly crucial for identifying accurate and relevant structural relationships between spatial events and factors that explain drivers’ behaviour while being classified and the explanations for them are evaluated. However, the high predictive power of ML algorithms ignore the characteristics of non-stationary domain relationships in spatiotemporal data (e.g., dependence, heterogeneity), which can lead to incorrect interpretations and poor management decisions. This study addresses this critical issue of ‘interpretability’ in ML-based modelling of structural relationships between the events and corresponding features of the car drivers’ behavioural variations. In this work, an exploratory experiment is described that contains simulator and real driving concurrently with a goal to enhance the simulator technologies. Here, initially, with heterogeneous data, several analytic techniques for simulator bias in drivers’ behaviour have been explored. Afterwards, five different ML classifier models were developed to classify risk and hurry in drivers’ behaviour in real and simulator driving. Furthermore, two different feature attribution-based explanation models were developed to explain the decision from the classifiers. According to the results and observation, among the classifiers, Gradient Boosted Decision Trees performed best with a classification accuracy of 98.62%. After quantitative evaluation, among the feature attribution methods, the explanation from Shapley Additive Explanations (SHAP) was found to be more accurate. The use of different metrics for evaluating explanation methods and their outcome lay the path toward further research in enhancing the feature attribution methods.
Download
Paper Nr: 272
Title:

On Converting Logic Programs Into Matrices
Authors:

Tuan N. Quoc and Katsumi Inoue
Abstract: Recently it has been demonstrated that deductive and abductive reasoning can be performed by exploiting the linear algebraic characterization of logic programs. Those experimental results reported so far on both forms of reasoning have proved that the linear algebraic approach can reach higher scalability than symbol manipulations. The main idea behind these proposed algorithms is based on linear algebra matrix multiplication. However, it has not been discussed in detail yet how to generate the matrix representation from a logic program in an efficient way. As conversion time and resulting matrix dimension are important factors that affect algebraic methods, it is worth investigating in standardization and matrix constructing steps. With the goal to strengthen the foundation of linear algebraic computation of logic programs, in this paper, we will clarify these steps and propose an efficient algorithm with empirical verification.
Download
Paper Nr: 273
Title:

JoVe-FL: A Joint-Embedding Vertical Federated Learning Framework
Authors:

Maria Hartmann, Grégoire Danoy, Mohammed Alswaitti and Pascal Bouvry
Abstract: Federated learning is a particular type of distributed machine learning, designed to permit the joint training of a single machine learning model by multiple participants that each possess a local dataset. A characteristic feature of federated learning strategies is the avoidance of any disclosure of client data to other participants of the learning scheme. While a wealth of well-performing solutions for different scenarios exists for Horizontal Federated Learning (HFL), to date little attention has been devoted to Vertical Federated Learning (VFL). Existing approaches are limited to narrow application scenarios where few clients participate, privacy is a main concern and the vertical distribution of client data is well-understood. In this article, we first argue that VFL is naturally applicable to another, much broader application context where sharing of data is mainly limited by technological instead of privacy constraints, such as in sensor networks or satellite swarms. A VFL scheme applied to such a setting could unlock previously inaccessible on-device machine learning potential. We then propose the Joint-embedding Vertical Federated Learning framework (JoVe-FL), a first VFL framework designed for such settings. JoVe-FL is based on the idea of transforming the vertical federated learning problem to a horizontal one by learning a joint embedding space, allowing us to leverage existing HFL solutions. Finally, we empirically demonstrate the feasibility of the approach on instances consisting of different partitionings of the CIFAR10 dataset.
Download
Paper Nr: 290
Title:

Latent Code Disentanglement Using Orthogonal Latent Codes and Inter-Domain Signal Transformation
Authors:

Babak Solhjoo and Emanuele Rodolà
Abstract: Auto Encoders are specific types of Deep Neural Networks that extract latent codes in a lower dimensional space for the inputs that are expressed in the higher dimensions. These latent codes are extracted by forcing the network to generate similar outputs to the inputs while limiting the data that can flow through the network in the latent space by choosing a lower dimensional space (Bank et al., 2020). Variational Auto Encoders realize a similar objective by generating a distribution of the latent codes instead of deterministic latent codes (Cosmo et al., 2020). This work focuses on generating semi-orthogonal variational latent codes for the inputs from different source types such as voice, image, and text for the same objects. The novelty of this work is on aiming to obtain unified variational latent codes for different manifestations of the same objects in the physical world using orthogonal latent codes. In order to achieve this objective, a specific Loss Function has been introduced to generate semi-orthogonal and variational latent codes for different objects. Then these orthogonal codes have also been exploited to map different manifestations of the same objects to each other. This work also uses these codes to convert the manifestations from one domain to another one.
Download
Paper Nr: 300
Title:

Generation-Based Data Augmentation Pipeline for Real-Time Automatic Gesture Recognition
Authors:

Mingxi Cheng, Fatima Z. Daha, Amit Srivastava and Ji Li
Abstract: With the SARS-CoV-2 pandemic outbreak, video conferencing tools experience huge spikes in usage. Gesture recognition can automatically translate non-verbal gestures into emoji reactions in these tools, making it easier for participants to express themselves. Nonetheless, certain rare gestures may trigger false alarms, and acquiring data for these negative classes in a timely manner is challenging. In this work, we develop a low-cost fast-to-market generation-based approach to effectively reduce the false alarm rate for any identified negative gesture. The proposed pipeline is comprised of data augmentation via generative adversarial networks, automatic gesture alignment, and model retraining with synthetic data. We evaluated our approach on a 3D-CNN based real-time gesture recognition system at a large software company. Experimental results demonstrate that the proposed approach can effectively reduce false alarm rate while maintaining similar accuracy on positive gestures.
Download
Paper Nr: 303
Title:

GaSUM: A Genetic Algorithm Wrapped BERT for Text Summarization
Authors:

Imen Tanfouri and Fethi Jarray
Abstract: Automatic Text Summarization (ATS) is a fundamental problem in natural language processing (NLP), it aims to reduce text size by removing irrelevant data and preserving the semantic structure of the original text. Recently, transformer-based models have shown great success in ATS and have been considered the state-of-the-art model for many NLP tasks. In this research, we are concerned with extractive summarization for a single document, where the goal is to extract a subset of sentences that best represents a summary of the document. We propose a combination of Bidirectional Encoder Representations from Transformers (BERT) and a Genetic Algorithm (GA) for automatic text summarization (named GaSUM) where GA is used as a search space method and BERT is used as a fitness measure. We validated these methods using the CNN/Daily Mail available dataset. Our results showed that GaSUM achieves a ROUGE-1 score of 55.75% and outperforms the state-of-the-art methods by a significant margin in terms of the rouge score.
Download
Paper Nr: 308
Title:

Brain Tumor Segmentation of Lower-Grade Glioma Across MRI Images Using Hybrid Convolutional Neural Networks
Authors:

Amal Jlassi, Khaoula ElBedoui and Walid Barhoumi
Abstract: Low-Grade Gliomas (LGG) are the most common malignant brain tumors that greatly define the rate of survival of patients. LGG segmentation across Magnetic Resonance Imaging (MRI) is common and necessary for diagnosis and treatment planning. To achieve this challenging clinical need, a deep learning approach that combines Convolutional Neural Networks (CNN) based on the hybridization of U-Net and SegNet is developed in this study. In fact, an adopted SegNet model was established in order to compare it with the most used model U-Net. The segmentation uses FLuid Attenuated Inversion Recovery (FLAIR) of 110 patients of LGG for training and evaluations. The highest mean and median Dice Coefficient (DC) achieved by the hybrid model is 83% and 85:7%, respectively. The obtained results of this work lead to the potential of using deep learning in MRI images in order to provide a non-invasive tool for automated LGG segmentation for many relevant clinical applications.
Download
Paper Nr: 310
Title:

MRNN: A Multi-Resolution Neural Network with Duplex Attention for Deep Ad-Hoc Retrieval
Authors:

Tolgahan Cakaloglu, Xiaowei Xu and Roshith Raghavan
Abstract: The primary goal of ad-hoc retrieval is to find relevant documents satisfying the information need posted by a natural language query. It requires a good understanding of the query and all the documents in a corpus, which is difficult because the meaning of natural language texts depends on the context, syntax, and semantics. Recently deep neural networks have been used to rank search results in response to a query. In this paper, we devise a multi-resolution neural network (MRNN) to leverage the whole hierarchy of representations for ad-hoc retrieval. The proposed MRNN model is capable of deriving a representation that integrates representations of different levels of abstraction from all the layers of the learned hierarchical representation. Moreover, a duplex attention component is designed to refine the multi-resolution representation so that an optimal context for matching the query and document can be determined. More specifically the first attention mechanism determines optimal context from the learned multi-resolution representation for the query and document. The latter attention mechanism aims to fine-tune the representation so that the query and the relevant document are closer in proximity. The empirical study shows that MRNN with the duplex attention is significantly superior to existing models used for ad-hoc retrieval on benchmark datasets including SQuAD, WikiQA, QUASAR, and TrecQA.
Download
Paper Nr: 312
Title:

CSP-DC: Data Cleaning via Constraint Satisfaction Problem Solving
Authors:

Nibel Nadjeh, Sabrina Abdellaoui and Fahima Nader
Abstract: In this paper, we present CSP-DC, a data cleaning system that integrates a new intelligent solution into the cleaning process to improve data accuracy, consistency, and minimize user involvement. We address three main challenges: (1) Consistency: Most repairing algorithms introduce new violations when repairing data, especially when constraints have overlapping attributes. (2) Automaticity: User intervention is time-consuming, we seek to minimize their efforts. (3) Accuracy: Most automatic approaches compute minimal repairs and apply unverified modifications to repair ambiguous cases, which may introduce more noise. To address these challenges, we propose to formulate this problem as a constraint satisfaction problem (CSP) allowing updates that always maintain data consistency. To achieve high performances, we perform a first cleaning phase to automatically repair violations that are easily handled by existing repair algorithms. We handle violations with multiple possible repairs with a CSP solving algorithm, which selects from possible fixes, values that respect all constraints. To reduce the problem’s complexity, we propose a new variables ordering technique and pruning strategies, allowing to optimize the repair search and find a solution quickly. Our experiments show that CSP-DC provides consistent and accurate repairs in a linear time, while also minimizing user intervention.
Download
Paper Nr: 313
Title:

Integrating Unsupervised Clustering and Label-Specific Oversampling to Tackle Imbalanced Multi-Label Data
Authors:

Payel Sadhukhan, Arjun Pakrashi, Sarbani Palit and Brian M. Namee
Abstract: There is often a mixture of very frequent labels and very infrequent labels in multi-label datasets. This variation in label frequency, a type class imbalance, creates a significant challenge for building efficient multi-label classification algorithms. In this paper, we tackle this problem by proposing a minority class oversampling scheme, UCLSO, which integrates Unsupervised Clustering and Label-Specific data Oversampling. Clustering is performed to find out the key distinct and locally connected regions of a multi-label dataset (irrespective of the label information). Next, for each label, we explore the distributions of minority points in the cluster sets. Only the intra-cluster minority points are used to generate the synthetic minority points. Despite having the same cluster set across all labels, we will use the label-specific class information to obtain a variation in the distributions of the synthetic minority points (in congruence with the label-specific class memberships within the clusters) across the labels. The training dataset is augmented with the set of label-specific synthetic minority points, and classifiers are trained to predict the relevance of each label independently. Experiments using 12 multi-label datasets and several multi-label algorithms shows the competency of the proposed method over other competing algorithms in the given context.
Download
Paper Nr: 314
Title:

Analysing Customer Behaviour Using Simulated Transactional Data
Authors:

Ryan Butler and Edwin Simpson
Abstract: This paper explores a novel technique that can aid firms in ascertaining a customer’s risk profile for the purpose of safeguarding them from unsuitable financial products. This falls under the purview of Know Your Customer (KYC), and a significant amount of regulation binds firms to this standard, including the Financial Conduct Authority (FCA) handbook Section 5.2. We introduce a methodology for computing a customer’s risk score by converting their transactional data into a heatmap image, then extracting complex geometric features that are indicative of impulsive spending. This heatmap analysis provides an interpretable approach to analysing spending patterns. The model developed by this study achieved an F1 score of 94.6% when classifying these features, far outperforming alternative configurations. Our experiments used a transactional dataset produced by Lloyds Banking Group, a major UK retail bank, via agent-based modelling (ABM). This data was computer generated and at no point was real transactional data shared. This study shows that a combination of ABM and artificial intelligence techniques can be used to aid firms in adhering to financial regulation.
Download
Short Papers
Paper Nr: 4
Title:

Why Do We Need Domain-Experts for End-to-End Text Classification? An Overview
Authors:

Jakob S. Andersen
Abstract: The aim of this study is to provide an overview of human-in-the-loop text classification. Automated text classification faces several challenges that negatively affect its applicability in real-world domains. General obstacles are a lack of labelled examples, limited held-out accuracy, missing user trust, run-time constraints, low data quality and natural fuzziness. Human-in-the-loop is an emerging paradigm to continuously support machine processing, i.e. text classification, with prior human knowledge, aiming to overcome the limitations of purely artificial processing. In this survey, we review current challenges of pure automated text classifiers and outline how a human-in-the-loop can overcome these obstacles. We focus on end-to-end text classification and feedback of domain-experts, which do not process technical knowledge about the algorithms used. Further, we discuss common techniques to guide human attention and efforts within the text classification process.
Download
Paper Nr: 5
Title:

Towards Low-Budget Real-Time Active Learning for Text Classification via Proxy-Based Data Selection
Authors:

Jakob S. Andersen and Olaf Zukunft
Abstract: Training data is typically the bottleneck of supervised machine learning applications, heavily relying on cost-intensive human annotations. Active Learning proposes an interactive framework to efficiently spend human efforts in the training data generation process. However, re-training state-of-the-art text classifiers is highly computationally intensive, leading to long training cycles that cause annoying interruptions to humans in the loop. To enhance the applicability of Active Learning, we investigate low-budget real-time Active Learning via Proxy-based data selection in the domain of text classification. We aim to enable fast interactive cycles within a minimal labelling effort while exploiting the performance of state-of-the-art text classifiers. Our results show that Proxy-based Active Learning can increase the F1-score of a lightweight classifier compared to a traditional budget Active Learning approach up to ~19%. Our novel Proxy-based Active Learning approach can be carried out time-efficiently, requiring less than 1 second for each learning iteration.
Download
Paper Nr: 6
Title:

A New Approach to Probabilistic Knowledge-Based Decision Making
Authors:

Thomas C. Henderson, Tessa Nishida, Amelia C. Lessen, Nicola Wernecke, Kutay Eken and David Sacharny
Abstract: Autonomous agents interact with the world by processing percepts and taking actions in order to achieve a goal. We consider agents which account for uncertainty when evaluating the state of the world, determine a high level goal based on this analysis, and then select an appropriate plan to achieve that goal. Such knowledge-based agents must take into account facts which are always true (e.g., laws of nature or rules) and facts which have some amount of uncertainty. This leads to probabilistic logic agents which maintain a knowledge base of facts each with an associated probability. We have previously described NILS, a nonlinear systems approach to solving atom probabilities, and compare it here to a hand-coded probability algorithm and a Monte Carlo method based on sampling possible worlds. We provide experimental data comparing the performance of these approaches in terms of successful outcomes in playing Wumpus World. The major contribution is the demonstration that the NILS method performs better than the human coded algorithm and is comparable to the Monte Carlo method. This advances the state-of-the-art in that NILS has been shown to have super-quadratic convergence rates.
Download
Paper Nr: 9
Title:

Enhancing Time Series Classification with Self-Supervised Learning
Authors:

Ali Ismail-Fawaz, Maxime Devanne, Jonathan Weber and Germain Forestier
Abstract: Self-Supervised Learning (SSL) is a range of Machine Learning techniques having the objective to reduce the amount of labeled data required to train a model. In Deep Learning models, SSL is often implemented using specific loss functions relying on pretext tasks leveraging from unlabelled data. In this paper, we explore SSL for the specific task of Time Series Classification (TSC). In the last few years, dozens of Deep Learning architectures were proposed for TSC. However, they almost exclusively rely on the traditional training step involving only labeled data in sufficient numbers. In order to study the potential of SSL for TSC, we propose the TRIplet Loss In TimE (TRILITE) which relies on an existing triplet loss mechanism and which does not require labeled data. We explore two use cases. In the first one, we evaluate the interest of TRILITE to boost a supervised classifier when very few labeled data are available. In the second one, we study the use of TRILITE in the context of semi-supervised learning, when both labeled and unlabeled data are available. Experiments performed on 85 datasets from the UCR archive reveal interesting results on some datasets in both use cases.
Download
Paper Nr: 17
Title:

A System for Updating Trust and Performing Belief Revision
Authors:

Aaron Hunter and Sam Tadey
Abstract: The process of belief revision is impacted by trust. In particular, when new information is received, it is only believed if the source is trusted as an authority on the given information. Moreover, trust is actually developed over time based on the accuracy of past reports. Any practical tool for representing and reasoning about the beliefs of communicating agents will therefore require some mechanism for modeling trust, both in terms of how it changes over time and in terms of how it impacts belief revision. In this paper, we present such a tool. We use so-called trust graphs to give a compact representation of how strongly one agent trusts another to distinguish between possible states of the world. Our software allows a trust graph to be updated incrementally by looking at the accuracy of past reports. After constructing a trust graph, the software can then compute the result of AGM-style belief revision using two different approaches to incorporating trust. In the first approach, trust is treated as a binary notion where an agent is either trusted to distinguish certain states or they are not. In the second approach, the relative strength of trust is compared directly with the strength of the initial beliefs. The end result is a tool that can flexibly model and reason about the dynamics of trust and belief.
Download
Paper Nr: 22
Title:

Fully Hidden Dynamic Trigger Backdoor Attacks
Authors:

Shintaro Narisada, Seira Hidano and Kazuhide Fukushima
Abstract: Indistinguishable adversarial attacks have been demonstrated with the sophistication of adversarial machine learning for neural networks. One example of such advanced algorithms is the backdoor attack with hidden triggers proposed by Saha et al. While Saha’s backdoor attack can produce invisible and dynamic triggers during the training phase without mislabeling, visible patch images are appended during the inference phase. A natural question is whether there exists a clean label backdoor attack whose trigger is dynamic and invisible at all times. In this study, we answer this question by adapting Saha’s backdoor attack to the trigger generation algorithm and by presenting a completely invisible backdoor attack with dynamic triggers and correct labels. Experimental results show that our proposed algorithm outperforms Saha’s backdoor attacks in terms of both indistinguishability and the attack success rate. In addition, we realize that our backdoor attack is a generalization of adversarial examples since our algorithm also works by using poisoning data only during the inference phase. We also describe a concrete algorithm for reconstructing adversarial examples as clean-label backdoor attacks. Several defensive experiments are conducted for both algorithms. This paper discovers the close relationship between hidden trigger backdoor attacks and adversarial examples.
Download
Paper Nr: 23
Title:

On the Impact of Grounding on HTN Plan Verification via Parsing
Authors:

Simona Ondrčková, Roman Barták, Pascal Bercher and Gregor Behnke
Abstract: The problem of hierarchical plan verification focuses on checking whether an action sequence is a valid hierarchical plan – the action sequence is executable and a goal task can be decomposed into it. The existing parsing-based verifier works on lifted domain models. In this paper we study whether grounding of the models could improve efficiency of the verifier. We also explore additional implementation improvements to increase the speed of the verifier.
Download
Paper Nr: 27
Title:

Nearest Neighbours and XAI Based Approach for Soft Labelling
Authors:

Ramisetty Kavya, Jabez Christopher and Subhrakanta Panda
Abstract: Hard label assignment is a challenging task in case of epistemic uncertainty. This work initially converts the hard labels of evidential instances into probabilistic labels based on k-nearest neighbours. Neighbours are identified in a way that at least half of them must belong to the hard label of the corresponding evidential instance. The probabilistic label of a decision query is computed by combining the probabilistic labels of the nearest neighbours using Dempster’s combination rule. Synthetic data is considered to verify the probabilistic labels over hard labels by varying the number of samples, number of neighbours and the overlapping degree between the classes. It is observed that the performance of the method mainly depends on the overlapping degree between classes. Probabilistic labels are intuitive compared to hard labels in case of high overlapping region. Moreover, few publicly available datasets are also considered to verify the performance of probabilistic labels on boundary instances. The proposed method achieves an accuracy of 90:44% and 98:24% on breast dataset trained with 10% and 90% of data respectively. Therefore the proposed method is sample efficient, calibrated, and interpretable.
Download
Paper Nr: 32
Title:

XR4DRAMA Knowledge Graph: A Knowledge Graph for Media Planning
Authors:

Alexandros Vassiliades, Spyridon Symeonidis, Sotiris Diplaris, Georgios Tzanetis, Stefanos Vrochidis and Ioannis Kompatsiaris
Abstract: In the previous two decades, knowledge graphs have evolved, inspiring developers to build even more contextrelated Knowledge Graphs. Because of this development, artificial intelligence applications can now access open domain-specific information in a format that is both semantically rich and machine comprehensible. In this paper, we introduce the XR4DRAMA Knowledge Graph, which can serve as a representation of media planning information. The XR4DRAMA knowledge graph can specifically represent data about the following: (a) Observations and Events (for example, data information from photos and text messages); (b) Spatial and Temporal data, such as coordinates or labels of locations and timestamps; and (c) Tasks and Plans for media planning. In addition, we provide a mechanism that allows Points of Interest to be created or updated based on videos, photos, and text messages sent by users. For improved media coverage of a remote location, Points of Interest serve as markers to journalists.
Download
Paper Nr: 39
Title:

Features Normalisation and Standardisation (FNS): An Unsupervised Approach for Detecting Adversarial Attacks for Medical Images
Authors:

Sreenivasan Mohandas and Naresh Manwani
Abstract: Deep learning systems have shown state-of-the-art performance in clinical prediction tasks. However, current research suggests that cleverly produced hostile images can trick these systems. Deep learning-based medical image classification algorithms have been questioned regarding their practical deployment. To address this problem, we provide an unsupervised learning technique for detecting adversarial attacks on medical images. Without identifying the attackers or reducing classification performance, our suggested strategy FNS (Features Normalization and Standardization), can detect adversarial attacks more effectively than earlier methods.
Download
Paper Nr: 40
Title:

SiameseBERT: A Bert-Based Siamese Network Enhanced with a Soft Attention Mechanism for Arabic Semantic Textual Similarity
Authors:

Rakia Saidi, Fethi Jarray and Mohammed Alsuhaibani
Abstract: The assessment of semantic textual similarity (STS) is a challenging task in natural language processing. It is crucial for many applications, including question answering, plagiarism detection, machine translation, information retrieval, and word sense disambiguation. The STS task evaluates the similarity of data pairs of text. For high high-resource languages (e.g. English), several approaches for STS have been proposed. In this paper, we are interested in measuring the semantic similarity of texts for Arabic, a low-resource language. A standard approach for STS is based on vector embedding of the input text and application of similarity metric on space embedding. In this contribution, we propose a BERT-based Siamese Network (SiameseBERT) and investigate the most available Arabic BERT models to embed the input sentences. We validate our approach via Arabic STS datasets. The araBERT-based Siamese Network model achieves a Pearson correlation of 0.925. The results obtained demonstrate the superiority of integrating the BERT embedding, the attention mechanism, and the Siamese neural network for the semantic textual similarity task.
Download
Paper Nr: 44
Title:

A Clustering Strategy for Enhanced FL-Based Intrusion Detection in IoT Networks
Authors:

Jacopo Talpini, Fabio Sartori and Marco Savi
Abstract: The Internet of Things (IoT) is growing rapidly and so the need of ensuring protection against cybersecurity attacks to IoT devices. In this scenario, Intrusion Detection Systems (IDSs) play a crucial role and data-driven IDSs based on machine learning (ML) have recently attracted more and more interest by the research community. While conventional ML-based IDSs are based on a centralized architecture where IoT devices share their data with a central server for model training, we propose a novel approach that is based on federated learning (FL). However, conventional FL is ineffective in the considered scenario, due to the high statistical heterogeneity of data collected by IoT devices. To overcome this limitation, we propose a three-tier FL-based architecture where IoT devices are clustered together based on their statistical properties. Clustering decisions are taken by means of a novel entropy-based strategy, which helps improve model training performance. We tested our solution on the CIC-ToN-IoT dataset: our clustering strategy increases intrusion detection performance with respect to a conventional FL approach up to +17% in terms of F1-score, along with a significant reduction of the number of training rounds.
Download
Paper Nr: 45
Title:

Interpretability and Explainability of Logistic Regression Model for Breast Cancer Detection
Authors:

Emina Tahirović and Senka Krivić
Abstract: Artificial Intelligence techniques are widely used for medical purposes nowadays. One of the crucial applications is cancer detection. Due to the sensitivity of such applications, medical workers and patients interacting with the system must get a reliable, transparent, and explainable output. Therefore, this paper examines the interpretability and explainability of the Logistic Regression Model (LRM) for breast cancer detection. We analyze the accuracy and transparency of the LRM model. Additionally, we propose an NLP-based interface with a model interpretability summary and a contrastive explanation for users. Together with textual explanations, we provide a visual aid for medical practitioners to understand the decision-making process better.
Download
Paper Nr: 52
Title:

An Automatic Method for Building a Taxonomy of Areas of Expertise
Authors:

Thu T. Le, Tuan-Dung Cao, Lam X. Pham, Trung D. Pham and Toan Luu
Abstract: Although a lot of Expert finding systems have been proposed, there is a need for a comprehensive study on building a knowledge base of areas of expertise. Building an Ontology creates a consistent lexical framework of a domain for representing information, thus processing the data effectively. This study uses the background knowledge of machine learning methods and textual data mining techniques to build adaptive clustering, local embedding, and term ordering modules. By that means, it is possible to construct an Ontology for a domain via representation language and apply it to the Ontology system of expert information. We proposed a new method called TaxoGenDRK (Taxonomy Generator using Database about Research Area and Keyword) based on the method from Chao Zhang et al. (2018)’s research on TaxoGen and an additional module that uses a database of research areas and keywords retrieved from the internet – the data regarded as an uncertain knowledge base for learning about taxonomy. DBLP dataset was used for testing, and the topic was “computer science”. The evaluation of the topic taxonomy using TaxogenDRK was implemented via qualitative and quantitative methods, producing a relatively good accuracy compared to other existing studies.
Download
Paper Nr: 56
Title:

A Survey of Socio-Emotional Strategies for Generation-Based Conversational Agents
Authors:

Lorraine Vanel, Alya Yacoubi and Chloé Clavel
Abstract: As dialogue systems are expected to display more and more “human” features, a subfield of conversational artificial intelligence (AI) has emerged, aiming to make virtual agents socially competent. As the common outlook firmly places emotion in the domain of chit-chatting, most of the studies tackle the problem of social behaviour in open domain applications. In this paper, we provide an overview of such approaches and think about how they can be applied to a task-oriented setting. The angle we explore is the influence of data on socio-emotional neural generation models. We focus on three aspects: dialogue strategies, emotional strategies and persona.
Download
Paper Nr: 59
Title:

Automatic Detection of Implicit and Typical Implementation of Singleton Pattern Based on Supervised Machine Learning
Authors:

Abir Nacef, Sahbi Bahroun, Adel Khalfallah and Samir Ben Ahmed
Abstract: Reverse engineering, based on design pattern recognition and software architecture refactoring, allows practitioners to focus on the overall architecture of the system without worrying about the programming details used to implement it. In this paper, we focus on the automatization of these tasks working on Singleton design Pattern (SP). The first task is the detection of the SP in its standard form, we named the detected structure as Singleton Typical implementation (ST). The second task consists of detecting structures which need the injection of the SP (Refactoring), these structures are named Singleton Implicit implementations (SI). All SP detection methods can only recover the typical form, even if they support different variants. However, in this work, we propose an approach based on supervised Machine Learning (ML) to extract different variants of SP in both ST and SI implementations and filter out structures which are incoherent with the SP intent. Our work consists of three phases; the first phase includes SP analysis, identifying implementation variants (ST and SI), and defining features for identifying them. In the second phase, we will extract feature values from the Java program using the LSTM classifier based on structural and semantic analysis. LSTM is trained on specific created data named SFD for a classification task. The third phase is SP detection, we create an ML classifier based on different algorithms, the classifier is named SPD. For training the SPD we create a new structured data named SDD constructed from features combination values that identify each variant. The SPD reaches 97% in terms of standard measures and outperforms the state-of-the-art approaches on the DPB corpus.
Download
Paper Nr: 64
Title:

Head Star (H*): A Motion Planning Algorithm for Navigation Among Movable Obstacles
Authors:

Halim Djerroud
Abstract: The objective of Navigation Among Movable Obstacles (NAMO) is to optimise the behaviour of robots by giving them the ability to manipulate obstacles. Current NAMO methods use two planners, one for moving through open spaces and a second for handling obstacles. However, these methods focus on providing a solution for obstacles handling and neglect movement in free spaces. These methods usually assume using classical obstacle avoidance algorithms for moving in free spaces. However, they are not suitable for the NAMO. This paper proposes a new path planning algorithm Head Star (H*) adapted for NAMO in free spaces. It is inspired by Bug’s algorithms by adding a graphical representation and heuristics on the distances allowing it to bring the robot as close as possible to its goal while keeping in memory the areas already visited.
Download
Paper Nr: 71
Title:

Multi-Objective Deep Q-Networks for Domestic Hot Water Systems Control
Authors:

Mohamed-Harith Ibrahim, Stéphane Lecoeuche, Jacques Boonaert and Mireille Batton-Hubert
Abstract: Real-world decision problems, such as Domestic Hot Water (DHW) production, require the consideration of multiple, possibly conflicting objectives. This work suggests an adaptation of Deep Q-Networks (DQN) to solve multi-objective sequential decision problems using scalarization functions. The adaptation was applied to train multiple agents to control DHW systems in order to find possible trade-offs between comfort and energy cost reduction. Results have shown the possibility of finding multiple policies to meet preferences of different users. Trained agents were tested to ensure hot water production with variable energy prices (peak and off-peak tariffs) for several consumption patterns and they can reduce energy cost from 10.24 % without real impact on users’ comfort and up to 18 % with slight impact on comfort.
Download
Paper Nr: 73
Title:

Generation of Daily and Monthly Flows Using the GR4j Method with ERA5 Grilled Data in the Cañete River Basin to the Putinza Hydrometric Station
Authors:

Edgar I. Quispe, Gianfranco G. Buitrón, Abel C. Arteaga and Rubén M. Chipana
Abstract: The research is carried out in a sub-basin of the Cañete River, delimited from the Putinza hydrometric station, with the aim of being able to generate the flows at a daily and monthly rate during a period of 39 years (1980 - 2019) and determine the approximate values of maximum flows in the periods that the El Niño phenomenon existed in the aforementioned basin, the methodology used was the GR4j method. On the one hand, the ERA5 grid data belonging to the European Space Agency Satellite was used, using Google Earth Engine (GEE) from which precipitation and average temperature information was extracted. Likewise, from the National Water System (ANA), information was extracted on daily flows from the Putinza hydrometric station between the period 2014-2017, which was used for the calibration and validation of the model. The analysis of results was carried out taking into account the Nash coefficient and the coefficient of determination R2 as efficiency criteria. Finally, the results obtained in the calibration and validation are satisfactory, which indicates that there is a good performance.
Download
Paper Nr: 77
Title:

Predicting Visual Importance of Mobile UI Using Semantic Segmentation
Authors:

Ami Yamamoto, Yuichi Sei, Yasuyuki Tahara and Akihiko Ohsuga
Abstract: When designing a UI, it is necessary to understand what elements are perceived to be important to users. The UI design process involves iteratively improving the UI based on feedback and eye-tracking results on the UI created by the designer, but this iterative process is time-consuming and costly. To solve this problem, several studies have been conducted to predict the visual importance of various designs. However, no studies specifically focus on predicting the visual importance of mobile UI. Therefore, we propose a method to predict visual importance maps from mobile UI screenshot images and semantic segmentation images of UI elements using deep learning. The predicted visual importance maps were objectively evaluated and found to be higher than the baseline. By combining the features of the semantic segmentation images appropriately, the predicted map became smoother and more similar to the ground truth.
Download
Paper Nr: 79
Title:

NewsRecs: A Mobile App Framework for Conducting and Evaluating Online Experiments for News Recommender Systems
Authors:

Noah Janzen and Fatih Gedikli
Abstract: News recommender systems are ubiquitous on the web. Intensive research has been conducted over the last decades, resulting in the continuous proposal of new recommendation techniques based on Machine Learning models. To evaluate the performance of recommendation algorithms, offline experiments, user studies, and online experiments should ideally be carried out one after the other so that the candidates move through a quality funnel. However, our literature review of multiple academic papers shows that new models have generally been evaluated using offline experiments only. Presumably, this is because researchers rarely have access to a production system. This work attempts to alleviate this problem by presenting a framework that can be used to evaluate recommendation models for news articles in an online scenario. The framework consists of a mobile app in which users can receive recommendations from different algorithms depending on their assigned group and rate them in multiple ways. The backend collects log data and makes it available for the final evaluation. The specific contributions our article will make are as follows: (1) A thematic review of 27 academic experiments from the news recommendation domain focusing on the evaluation design. (2) An open-source mobile app framework for conducting and evaluating online experiments.
Download
Paper Nr: 80
Title:

Emotions Relationship Modeling in the Conversation-Level Sentiment Analysis
Authors:

Jieying Xue, Minh-Phuong Nguyen and Le-Minh Nguyen
Abstract: Sentiment analysis, also called opinion mining, is a task of Natural Language Processing (NLP) that aims to extract sentiments and opinions from texts. Among them, emotion recognition in conversation (ERC) is becoming increasingly popular as a new research topic in natural language processing (NLP). The current state-of-the-art models focus on injecting prior knowledge via an external commonsense extractor or applying pre-trained language models to construct the utterance vector representation that is fused with the surrounding context in a conversation. However, these architectures treat the emotional states as sequential inputs, thus omitting the strong relationship between emotional states of discontinuous utterances, especially in long conversations. To solve this problem, we propose a new architecture, Long-range dependencY emotionS Model (LYSM) to generalize the dependencies between emotional states using the self-attention mechanism, which reinforces the emotion vector representations in the conversational encoder. Our intuition is that the emotional states in a conversation can be influenced or transferred across speakers and sentences, independent of the length of the conversation. Our experimental results show that our proposed architecture improves the baseline model and achieves competitive performance with state-of-the-art methods on four well-known benchmark datasets in this domain: IEMOCAP, DailyDialog, Emory NLP, and MELD. Our code is available at https://github.com/phuongnm94/erc-sentiment.
Download
Paper Nr: 84
Title:

Background Image Editing with HyperStyle and Semantic Segmentation
Authors:

Syuusuke Ishihata, Ryohei Orihara, Yuichi Sei, Yasuyuki Tahara and Akihiko Ohsuga
Abstract: Recently, research has been conducted on applying StyleGAN to image editing tasks. Although the technique can be applied to editing background images, because they are more diverse than foreground images such as face images, specifying an object in background images to be edited is difficult. For example, because natural language instructions can be ambiguous, edited images become undesirable for the user. It is challenging to resolve style and content dependencies in image editing. In our study, we propose an editing method that adapts Style Transformer, the latest GAN inversion encoder approach, to HyperStyle by introducing semantic segmentation to maintain the reconstruction quality and separate the style and the content of the background image. The content is edited while keeping the original style by manipulating the coarse part of latent variables and the residual parameters obtained by HyperStyle, and the style is edited without changing the content by manipulating the medium and fine part of latent vectors as in the conventional StyleGAN. As a result, the qualitative evaluation confirms that our model enabled the editing of image content and style separately, and the quantitative evaluation validates that the reconstruction quality is comparable to the conventional method.
Download
Paper Nr: 91
Title:

Diverse Level Generation for Tile-Based Video Game using Generative Adversarial Networks from Few Samples
Authors:

Soichiro Takata, Yuichi Sei, Yasuyuki Tahara and Akihiko Ohsuga
Abstract: The procedural generation of levels in video games has been studied mainly to reduce the burden on producers. In recent years, methods based on deep learning have been attracting attention. In level generations with deep learning, GAN-based methods have achieved some success in tile-based video games, but the preparation of the dataset has been an issue. In this study, we investigate a method to acquire a model that can generate various levels by learning a GAN from only a small amount of data. It was confirmed that a greater variety and lower playability of levels can be generated than with conventional methods by quantitative evaluation of the levels generated by the proposed methods. In addition, the model learned by the proposed method can generate levels that reflect the objectives more strongly than the conventional method by using CMA-ES to search for latent variables.
Download
Paper Nr: 95
Title:

Symbolic Explanations for Multi-Label Classification
Authors:

Ryma Boumazouza, Fahima Cheikh-Alili, Bertrand Mazure and Karim Tabia
Abstract: This paper proposes an agnostic and declarative approach to provide different types of symbolic explanations for multi-label classifiers. More precisely, in addition to global sufficient reason and counterfactual explanations, our approach makes it possible to generate explanations at different levels of granularity in addition to structural relationships between labels. Our approach is declarative and allows to take advantage of the strengths of modern SAT-based oracles and solvers. Our experimental study provides promising results on many multi-label datasets.
Download
Paper Nr: 100
Title:

Compressing UNSAT Search Trees with Caching
Authors:

Anthony Blomme, Daniel Le Berre, Anne Parrain and Olivier Roussel
Abstract: In order to provide users of SAT solvers with small, easily understandable proofs of unsatisfiability, we present caching techniques to identify redundant subproofs and reduce the size of some UNSAT proof trees. In a search tree, we prune branches corresponding to subformulas that were proved unsatisfiable earlier in the tree. To do so, we use a cache inspired by model counters and we adapt it to the case of unsatisfiable formulas. The implementation of this cache in a CDCL and a DPLL solver is discussed. This approach can drastically reduce the UNSAT proof tree of several benchmarks from the SAT’02 and SAT’03 competitions.
Download
Paper Nr: 107
Title:

An Effective Hybrid Text-Based Approach to Identify Fake News on Social Media
Authors:

Imtiez Fliss and Hamza Bargougui
Abstract: Because of their low cost, simplicity of access, and quick dissemination, social media are today one of the primary information sources for millions of people worldwide. However, this is at the expense of dubious credibility and a large danger of being exposed to ”fake news,” which is deliberately designed to mislead readers. In light of this, in this paper we propose a novel method for identifying bogus news based on the text content. This method is founded on a mix of BERT (Bidirectional Encoder Representations from Transformers) and deep learning techniques (Convolutional Neural Networks (CNN) and Long Short Term Memory (LSTM)). Promising results are seen when the proposed approach is compared to other models.
Download
Paper Nr: 109
Title:

Cost-Aware Ensemble Learning Approach for Overcoming Noise in Labeled Data
Authors:

Abdulrahman Gharawi, Jumana Alsubhi and Lakshmish Ramaswamy
Abstract: Machine learning models have demonstrated exceptional performance in various applications as a result of the emergence of large labeled datasets. Although there are many available datasets, acquiring high-quality labeled datasets is challenging since it involves huge human supervision or expert annotation, which are extremely labor-intensive and time-consuming. Since noisy datasets can affect the performance of machine learning models, acquiring high-quality datasets without label noise becomes a critical problem. However, it is challenging to significantly decrease label noise in real-world datasets without hiring expensive expert annotators. Based on extensive testing and research, this study examines the impact of different levels of label noise on the accuracy of machine learning models. It also investigates ways to cut labeling expenses without sacrificing required accuracy.
Download
Paper Nr: 113
Title:

GAN Inversion with Editable StyleMap
Authors:

So Honda, Ryohei Orihara, Yuichi Sei, Yasuyuki Tahara and Akihiko Ohsuga
Abstract: Recently, the field of GAN Inversion, which estimates the latent code of a GAN to reproduce the desired image, has attracted much attention. Once a latent variable that reproduces the input image is obtained, the image can be edited by manipulating the latent code. However, it is known that there is a trade-off between reconstruction quality, which is the difference between the input image and the reproduced image, and editability, which is the plausibility of the edited image. In our study, we attempted to improve reconstruction quality by extending latent code that represents the properties of the entire image in the spatial direction. Next, since such an expansion significantly impairs the editing quality, we performed a GAN Inversion that realizes both reconstruction quality and editability by imposing an additional regularization. As a result, the proposed method yielded a better trade-off between the reconstruction quality and the editability against the baseline from both quantitative and qualitative perspectives, and is comparable to state-of-the-art(SOTA) methods that adjust the weights of the generators.
Download
Paper Nr: 121
Title:

Explainable Decision Support Modelling Based on Multi-Layer FCM with Multi-Objective Optimization Characteristics: The Case of the Microservices Adoption Problem
Authors:

Andreas Christoforou and Andreas S. Andreou
Abstract: The tremendous progress in the field of artificial and computational intelligence has enabled the application of relevant techniques to a wide range of human life aspects. However, these techniques appear in their majority incompetent to allow users to explain and understand their decisions. This paper introduces an enhanced, explainable decision support approach using a promising graph-based computational intelligent model, namely Multi-Layer Fuzzy Cognitive Maps (MLFCM). MLFCM have evolved over the last two decades into a flexible and powerful tool that enables the execution of simulation scenarios to facilitate decision support in highly complex environments. The proposed enhancement of MLFCM revolves around their integration with MultiObjective Evolutionary Algorithms that allows executing simulations with multiple conflicting targets and then analyzing the values and relationships of the participating nodes. The applicability of the enhanced MLFCM is demonstrated through a case-study on adopting microservices. Microservices have been considered as one of the most promising alternatives to software development nowadays. Nevertheless, their adoption often stumbles on various factors such as security, exit policy, effectiveness, etc. In this context, the factors contributing to Microservices adoption are assessed, analyzed and modeled via MLFCM using a series of real-world and synthetic scenarios that yielded quite promising results.
Download
Paper Nr: 126
Title:

Thompson Sampling on Asymmetric a-stable Bandits
Authors:

Zhendong Shi, Ercan E. Kuruoglu and Xiaoli Wei
Abstract: In algorithm optimization in reinforcement learning, how to deal with the exploration-exploitation dilemma is particularly important. Multi-armed bandit problem can be designed to realize the dynamic balance between exploration and exploitation by changing the reward distribution. Thompson Sampling has been proposed in the literature for the solution of the multi-armed bandit problem by sampling rewards from posterior distributions. Recently, it was used to process non-Gaussian data with heavy tailed distributions. It is a common observation that various real-life data such as social network data and financial data demonstrate not only impulsive but also asymmetric characteristics. In this paper, we consider the Thompson Sampling approach for multi-armed bandit problem, in which rewards conform to an asymmetric a-stable distribution with unknown parameters and explore their applications in modelling financial and recommendation system data.
Download
Paper Nr: 128
Title:

Voicemail Urgency Detection Using Context Dependent and Independent NLP Techniques
Authors:

Asma Trabelsi, Séverine Soussilane and Emmanuel Helbert
Abstract: Business field has improved exponentially during the last two decades: working methods have changed, more and more users are connected to each other across the globe, same teams as well as different teams can be separated by countries in big companies. So, users need a way to select messages to treat in priority for a better business management and a better communication. In this paper, we implement an approach enabling to classify voicemail messages into urgent and non urgent. The problem of determining urgency being still vast and open, some criteria should be used to decide the importance of messages depending to one’s necessity. Among these criteria, we can mention the sender position, the time of sending as well as the textual content. In this paper, we focus on classifying voicemail messages based on their contents. As there exist several Machine Learning approaches for text vectorization and classification, various combinations will be discussed and compared for the aim of finding the most performant one.
Download
Paper Nr: 135
Title:

Evolving Behavioural Level Sequence Detectors in SystemVerilog Using Grammatical Evolution
Authors:

Bilal Majeed, Conor Ryan, Jack McEllin, Ayman Youssef, Douglas M. Dias, Aidan Murphy and Samuel Carvalho
Abstract: Sequential circuits are time-dependent circuits whose output depends not only on their current inputs but also on previous ones. This makes them substantially more complex than combinational circuits, which are stateless and only produce outputs from their current inputs. This paper demonstrates the automatic evolution of some of the most critical and hard-to-evolve electronic sequential circuits, namely, sequence detectors. The circuits are generated at behavioural level using the Hardware Description Language, SystemVerilog. We successfully evolve solutions ranging in complexity from 3 to 5 bits, with and without encapsulation, and 6 bits with encapsulation while using Grammatical Evolution. A uniform distribution of values that a vector of 50 bits can represent was used to generate the random training and test data sets to prevent any bias in the solutions and results. While previous work combined shorter sequence detectors to produce longer ones, for example, combining two 3-bit detectors to form a 6-bit detector, we produce all sequence detectors from scratch without any intermediate stages. The system simply takes instructions and testcases and produces the desired detector; we show that not only does it produce longer-sequence detectors than previous work, but it also does it using fewer computational resources.
Download
Paper Nr: 136
Title:

Multi-Agent Parking Problem with Sequential Allocation
Authors:

Aniello Murano, Silvia Stranieri and Munyque Mittelmann
Abstract: In this paper, we study the multi-agent parking problem with time constraints adopting a game-theoretic perspective. Precisely, cars are modeled as agents interacting among themselves in a multi-player game setting, each of which aims to find a free parking slot that satisfies their constraints. We provide an algorithm for assigning parking slots based on a sequential allocation with priorities. We show that the algorithm always finds a Nash equilibrium solution and we prove its complexity is in quadratic time. The usefulness of our approach is demonstrated by considering its application to the parking area of the Federico II Hospital Company in Naples. Finally, we provide experimental results comparing our algorithm with a greedy allocation and evaluating its performance in the application scenario.
Download
Paper Nr: 139
Title:

Semantically Layered Representation for Planning Problems and Its Usage for Heuristic Computation Using Cellular Simultaneous Recurrent Neural Networks
Authors:

Michaela Urbanovská and Antonín Komenda
Abstract: Learning heuristic functions for classical planning algorithms has been a great challenge in the past years. The biggest bottleneck of this technique is the choice of an appropriate description of the planning problem suitable for machine learning. Various approaches were recently suggested in the literature, namely grid-based, image-like, and graph-based. In this work, we extend the latest grid-based representation with layered architecture capturing the semantics of the related planning problem. Such an approach can be used as a domain-independent model for further heuristic learning. This representation keeps the advantages of the grid-structured input and provides further semantics about the problem we can learn from. Together with the representation, we also propose a new network architecture based on the Cellular Simultaneous Recurrent Networks (CSRN) that is capable of learning from such data and can be used instead of a heuristic function in the state-space search algorithms. We show how to model different problem domains using the proposed representation as well as explain the new neural network architecture and compare its performance in the state-space search against existing classical planning heuristics and heuristics provided by the state-of-the-art.
Download
Paper Nr: 145
Title:

Targeted Adversarial Attacks on Deep Reinforcement Learning Policies via Model Checking
Authors:

Dennis Gross, Thiago D. Simão, Nils Jansen and Guillermo A. Pérez
Abstract: Deep Reinforcement Learning (DRL) agents are susceptible to adversarial noise in their observations that can mislead their policies and decrease their performance. However, an adversary may be interested not only in decreasing the reward, but also in modifying specific temporal logic properties of the policy. This paper presents a metric that measures the exact impact of adversarial attacks against such properties. We use this metric to craft optimal adversarial attacks. Furthermore, we introduce a model checking method that allows us to verify the robustness of RL policies against adversarial attacks. Our empirical analysis confirms (1) the quality of our metric to craft adversarial attacks against temporal logic properties, and (2) that we are able to concisely assess a system’s robustness against attacks.
Download
Paper Nr: 148
Title:

Automatically Generating Image Segmentation Datasets for Video Games
Authors:

David G. LeBlanc and Greg Lee
Abstract: Image segmentation is applied to images fed as input to deep reinforcement learning agents as a way of highlighting key-features and removing non-key features. If a segmented image is of lower resolution than its source, the problem is further simplified. However, the process of creating a dataset for the training of an image segmenting network is long and costly if done manually. This paper proposes a methodology for automatically generating an arbitrarily large image segmentation dataset with a specifiable segmentation resolution. A convolutional neural network trained for image segmentation using this automatically generated dataset had higher accuracy than a network using a manually labelled training set. Furthermore, an image segmenting network trained on a dataset generated in this manner gave superior performance to an autoencoder in reducing dimensionality while preserving key features. The method proposed was tested on Super Mario Bros. for the Nintendo Entertainment System (NES), but the techniques could apply to any image segmentation problem where it is possible to simulate the placement of key objects.
Download
Paper Nr: 149
Title:

Candidate Path Selection Heuristics for Multi-Agent Path Finding: A Novel Compilation-Based Method
Authors:

Pavel Surynek
Abstract: Multi-Agent path finding (MAPF) is a task of finding non-conflicting paths connecting agents’ specified initial and goal positions in a shared environment. We focus on compilation-based solvers in which the MAPF problem is expressed in a different well established formalism such as mixed-integer linear programming (MILP), Boolean satisfiability (SAT), or constraint programming (CP). As the target solvers for these formalisms act as black-boxes it is challenging to integrate MAPF specific heuristics in the MAPF compilation-based solvers. We show in this work how the build a MAPF encoding for the target SAT solver in which domain specific heuristic knowledge is reflected.
Download
Paper Nr: 154
Title:

Video Game Agents with Human-like Behavior using the Deep Q-Network and Biological Constraints
Authors:

Takahiro Morita and Hiroshi Hosobe
Abstract: Video game agents that surpass humans always select the optimal behavior, which may make them look mechanical and uninteresting to human players and audience. Since score-oriented game agents have been almost achieved, a next goal should be to entertain human players and audience by realizing agents that reproduce human-like behavior. A previous method implemented such game agents by introducing biological constraints into Q-learning and A* search. In this paper, we propose video game agents with more entertaining and more practical human-like behavior by applying biological constraints into the deep Q-network (DQN). Especially, to reduce the problem of the conspicuous mechanical behavior found in the previous method, we propose an additional biological constraint “confusion”. We implemented our method in the video game “Infinite Mario Bros.” and conducted a subjective evaluation. The results indicated that the agents implemented with our method were rated more human-like than those implemented with the previous method.
Download
Paper Nr: 156
Title:

Learning Spatio-Temporal Features via 3D CNNs to Forecast Time-to-Accident
Authors:

Taif Anjum, Louis Chirade, Beiyu Lin and Apurva Narayan
Abstract: Globally, traffic accidents are one of the leading causes of death. Collision avoidance systems can play a critical role in preventing accidents or minimizing their severity. Time-to-accident (TTA) is considered the principal parameter for collision avoidance systems allowing for decision-making in traffic, dynamic path planning, and accident mitigation. Despite the importance of TTA, the literature has insufficient research on TTA estimation for traffic scenarios. The majority of recent work focuses on accident anticipation by providing a probabilistic measure of an immediate or future collision. We propose a novel approach of time-to-accident forecasting by predicting the exact time of the accident with a prediction horizon of 3-6 seconds. Leveraging the Spatio-temporal features from traffic accident videos, we can recognize accident and non-accident scenes while forecasting the TTA. Our method is solely image-based, using video data from inexpensive dashboard cameras allowing for an accessible collision avoidance tool that can be integrated with any vehicle. Additionally, we present a regression-based 3D Convolutional Neural Network (CNN) architecture that requires significantly less parameters compared to its counterparts making it feasible for real-time usage. Our best models can estimate TTA with an average prediction error of 0.30s on the Car Crash Dataset (CCD) and 0.79s on the Detection of Traffic Anomalies (DoTA) dataset elucidated by the longer prediction horizon. Our comprehensive experiments suggest that spatio-temporal features from sequential frames perform significantly better than only spatial features extracted from static images.
Download
Paper Nr: 163
Title:

InFER: A Multi-Ethnic Indian Facial Expression Recognition Dataset
Authors:

Syed A. Rizvi, Preyansh Agrawal, Jagat S. Challa and Pratik Narang
Abstract: The rapid advancement in deep learning over the past decade has transformed Facial Expression Recognition (FER) systems, as newer methods have been proposed that outperform the existing traditional handcrafted techniques. However, such a supervised learning approach requires a sufficiently large training dataset covering all the possible scenarios. And since most people exhibit facial expressions based upon their age group, gender, and ethnicity, a diverse facial expression dataset is needed. This becomes even more crucial while developing a FER system for the Indian subcontinent, which comprises of a diverse multi-ethnic population. In this work, we present InFER, a real-world multi-ethnic Indian Facial Expression Recognition dataset consisting of 10,200 images and 4,200 short videos of seven basic facial expressions. The dataset has posed expressions of 600 human subjects, and spontaneous/acted expressions of 6000 images crowd-sourced from the internet. To the best of our knowledge InFER is the first of its kind consisting of images from 600 subjects from very diverse ethnicity of the Indian Subcontinent. We also present the experimental results of baseline & deep FER methods on our dataset to substantiate its usability in real-world practical applications.
Download
Paper Nr: 170
Title:

Hybrid Genetic U-Net Algorithm for Medical Segmentation
Authors:

Jon-Olav Holland, Youcef Djenouri, Roufaida Laidi and Anis Yazidi
Abstract: U-Net based architecture has become the de-facto standard approach for medical image segmentation in recent years. Many researchers have used the original U-Net as a skeleton for suggesting more advanced models such as UNet++ and UNet 3+. This paper seeks to boost the performance of the original U-Net via optimizing its hyperparameters. Rather than changing the architecture itself, we optimize hyperparameters which does not affect the architecture, but affects the performance of the model. For this purpose, we use genetic algorithms. Intensive experiments on medical dataset have been carried out which document a performance gain at a low computation cost. In addition, preliminary results reveal the benefit of the proposed framework for medical image segmentation.
Download
Paper Nr: 173
Title:

Ethical Considerations for the Deployment of Logic-Based Models of Reasoning
Authors:

Aaron Hunter
Abstract: Ethical considerations in the development of Artificial Intelligence systems need to be addressed as more systems are deployed in practice. Much of the current work in this area is focused on Machine Learning systems, with an emphasis on issues such as fairness and bias. However, there are also fundamental ethical problems to be addressed in simple logic-based systems, and we do not have solid methods in place to handle these issues. In this paper, we discuss ethical problems that are implicitly introduced in the deployment of systems that formalize reasoning in logic. As a specific example, we focus on logic-based models of belief change. We consider the way belief change operators are defined, and how unintended behaviour can emerge in operators defined with respect to well-known rationality postulates. Preventative measures and potential solutions are discussed.
Download
Paper Nr: 174
Title:

Bagged Ensembles for Blood Glucose Prediction: A Comparative Study
Authors:

Mohamed Z. Wadghiri and Ali Idri
Abstract: Blood Glucose (BG) prediction is an essential process for diabetes self-management. Many papers investigated the use of various machine learning techniques to design and implement BGL predictors. However, due to the complexity of glucose dynamics, single techniques do not always capture inter- and intra-patient changes. On the other hand, ensemble learning and bagging ensembles in particular have been established to show better performance in many medical disciplines including diabetology. The aim of the present paper is to build BG predictors based on bagging in order to compare their performance to the accuracy of their underlying single techniques and to verify if a particular ensemble outperforms the others. An approach has been proposed to build bagged predictors based on five techniques: LSTM, GRU, CNN, SVR and DT. The models’ performance has been evaluated and compared at a prediction horizon of 30 minutes according to RMSE and CEGA. The results show that the performance of the constructed bagging ensembles is very comparable to their underlying single techniques except for regression trees. This can be attributed to the good accuracy of deep learning models but also to the non-stationarity of BG time series that need to be addressed before constructing the bootstrap samples.
Download
Paper Nr: 175
Title:

Shrinking the Inductive Programming Search Space with Instruction Subsets
Authors:

Edward McDaid and Sarah McDaid
Abstract: Inductive programming frequently relies on some form of search in order to identify candidate solutions. However, the size of the search space limits the use of inductive programming to the production of relatively small programs. If we could somehow correctly predict the subset of instructions required for a given problem then inductive programming would be more tractable. We will show that this can be achieved in a high percentage of cases. This paper presents a novel model of programming language instruction co-occurrence that was built to support search space partitioning in the Zoea distributed inductive programming system. This consists of a collection of intersecting instruction subsets derived from a large sample of open source code. Using the approach different parts of the search space can be explored in parallel. The number of subsets required does not grow linearly with the quantity of code used to produce them and a manageable number of subsets is sufficient to cover a high percentage of unseen code. This approach also significantly reduces the overall size of the search space - often by many orders of magnitude.
Download
Paper Nr: 177
Title:

How Fine Tuning Affects Contextual Embeddings: A Negative Result Explanation
Authors:

Ha-Thanh Nguyen, Vu Tran, Minh-Phuong Nguyen, Le-Minh Nguyen and Ken Satoh
Abstract: Recently, deep learning models trained on large amounts of data have achieved impressive results in the field of legal document processing. However, being seen as black boxes, these models lack explainability. This paper aims to shed light on the inner behavior of legal learning models by analyzing the effect of fine-tuning on legal contextual embeddings. This paper provides pieces of evidence to explain the relationship between the moving of contextual embeddings and the effectiveness of a model when fine-tuned on legal tasks. It can help further explain the effect of finetuning on language models. To this end, we use multilingual transformer models, fine-tune them on the lawfulness classification task, and record the changes in the embeddings. The experimental results reveal interesting phenomena. The method in this paper can be used to confirm whether a deep learning model truly gains the knowledge in a legal problem to make the predictions or simply memorize the training examples, or worse, predict randomly.
Download
Paper Nr: 178
Title:

Integration of Efficient Deep Q-Network Techniques Into QT-Opt Reinforcement Learning Structure
Authors:

Shudao Wei, Chenxing Li, Jan Seyler and Shahram Eivazi
Abstract: There has been a growing interest in the development of offline reinforcement learning (RL) algorithms for real-world applications. For example, offline algorithms like qt-opt has demonstrated an impressive performance in grasping task. The primary motivation is to avoid the challenges associated with online data collection. However, these algorithms require extremely large dataset as well as huge computational resources. In this paper we investigate the applicability of well known improvement techniques from Deep Q-learning (DQN) methods to the QT-Opt offline algorithm, for both on-policy and mixed-policy training. For the first time, we show that prioritized experience replay(PER), noisy network, and distributional DQN can be used within QT-Opt framework. As result,for example, in a reacher environment from Pybullet simulation, we observe an obvious improvements in the learning process for the integrated techniques.
Download
Paper Nr: 183
Title:

Modeling Syntactic Knowledge With Neuro-Symbolic Computation
Authors:

Hilton Alers-Valentín, Sandiway Fong and J. F. Vega-Riveros
Abstract: To overcome the limitations of prevailing NLP methods, a Hybrid-Architecture Symbolic Parser and Neural Lexicon system is proposed to detect structural ambiguity by producing as many syntactic representations as there are interpretations for an utterance. HASPNeL comprises a symbolic AI, feature-unification parser, a lexicon generated using manual classification and machine learning, and a neural network encoder which tags each lexical item in a synthetic corpus and estimates likelihoods for each utterance’s interpretation with respect to the corpus. Language variation is accounted for by lexical adjustments in feature specifications and minimal parameter settings. Contrary to pure probabilistic system, HASPNeL’s neuro-symbolic architecture will perform grammaticality judgements of utterances that do not correspond to rankings of probabilistic systems; have a greater degree of system stability as it is not susceptible to perturbations in the training data; detect lexical and structural ambiguity by producing all possible grammatical representations regardless of their presence in the training data; eliminate the effects of diminishing returns, as it does not require massive amounts of annotated data, unavailable for underrepresented languages; avoid overparameterization and potential overfitting; test current syntactic theory by implementing a Minimalist grammar formalism; and model human language competence by satisfying conditions of learnability, evolvability, and universality.
Download
Paper Nr: 184
Title:

Applied Deep Learning Architectures for Breast Cancer Screening Classification
Authors:

Asma Zizaan, Ali Idri and Hasnae Zerouaoui
Abstract: Breast cancer (BC) became the most diagnosed cancer, making it one of the deadliest diseases. Mammography is a modality used for early detection of breast cancer. The objective of the present paper is to evaluate and compare deep learning techniques applied to mammogram images. The paper conducts an experimental evaluation of eight deep Convolutional Neural Network (CNN) architectures for a binary classification of breast screening mammograms, namely VGG16, VGG19, DenseNet201, Inception ResNet V2, Inception V3, ResNet 50, MobileNet V2 and Xception. This evaluation was based on four performance metrics (accuracy, precision, recall and f1-score), as well as Scott Knott statistical test and Borda count voting system. The data was extracted from the CBIS-DDSM dataset with 4000 images. And results have shown that DenseNet201 was the most efficient model for the binary classification with an accuracy of 84.27%.
Download
Paper Nr: 185
Title:

Can We Use Probing to Better Understand Fine-Tuning and Knowledge Distillation of the BERT NLU?
Authors:

Jakub Hościłowicz, Marcin Sowański, Piotr Czubowski and Artur Janicki
Abstract: In this article, we use probing to investigate phenomena that occur during fine-tuning and knowledge distillation of a BERT-based natural language understanding (NLU) model. Our ultimate purpose was to use probing to better understand practical production problems and consequently to build better NLU models. We designed experiments to see how fine-tuning changes the linguistic capabilities of BERT, what the optimal size of the fine-tuning dataset is, and what amount of information is contained in a distilled NLU based on a tiny Transformer. The results of the experiments show that the probing paradigm in its current form is not well suited to answer such questions. Structural, Edge and Conditional probes do not take into account how easy it is to decode probed information. Consequently, we conclude that quantification of information decodability is critical for many practical applications of the probing paradigm.
Download
Paper Nr: 187
Title:

Evaluating and Improving End-to-End Systems for Knowledge Base Population
Authors:

Maxime Prieur, Cédric D. Mouza, Guillaume Gadek and Bruno Grilheres
Abstract: Knowledge Bases (KB) are used in many fields, such as business intelligence or user assistance. They aggregate knowledge that can be exploited by computers to help decision making by providing better visualization or predicting new relations. However, their building remains complex for an expert who has to extract and link each new information. In this paper, we describe an entity-centric method for evaluating an end-to-end Knowledge Base Population system. This evaluation is applied to ELROND, a complete system designed as a workflow composed of 4 modules (Named Entity Recognition, Coreference Resolution, Relation Extraction and Entity Linking) and MERIT, a dynamic entity linking model made of a textual encoder to retrieve similar entities and a classifier.
Download
Paper Nr: 189
Title:

Model Transparency: Why Do We Care?
Authors:

Ioannis Papantonis and Vaishak Belle
Abstract: Artificial intelligence (AI) and especially machine learning (ML) has been increasingly incorporated into a wide range of critical applications, such as healthcare, justice, credit risk assessment, and loan approval. In this paper, we survey the motivations for caring about model transparency, especially as AI systems are becoming increasingly complex leviathans with many moving parts. We then briefly outline the challenges in providing computational solutions to transparency.
Download
Paper Nr: 197
Title:

Bottom-up Japanese Word Ordering Using BERT
Authors:

Masato Yamazoe, Tomohiro Ohno and Shigeki Matsubara
Abstract: Although Japanese is widely regarded as a relatively free word order language, word order in Japanese is not entirely arbitrary and has some sort of preference. As a result, it is an important technique to produce a sentence that is not only grammatically correct but also easy to read. This paper proposes a method for the word ordering of a whole Japanese sentence when the dependency relations between words are given. Using BERT, the method identifies the easy-to-read word order for a syntactic tree based on bottom-up processing. We confirmed the effectiveness of our method through an experiment on word ordering using newspaper articles.
Download
Paper Nr: 200
Title:

Explainability of MLP Based Species Distribution Models: A Case Study
Authors:

Imane El Assari, Hajar Hakkoum and Ali Idri
Abstract: Species Distribution models (SDMs) are widely used to study species occurrence in conservation science and ecology evolution. However the huge amount of data and its complexity makes it difficult for professionals to forecast the evolutionary trends of distributions across the concerned landscapes. As a solution, machine learning (ML) algorithms were used to construct and evaluate SDMs in order to predict the studied species occurrences and their habitat suitability. Nevertheless, it is critical to ensure that ML based SDMs reflect reality by studying their trustworthiness. This paper aims to investigate two techniques: SHapley Additive exPlanations (SHAP) and the Partial Dependence Plot (PDP) techniques to interpret a Multilayer perceptron (MLP) trained on the Loxodonta Africana dataset. Results demonstrate the prediction process and how in- terpretability techniques could be used to explain misclassified instances and thus increase trust between ML results and domain experts.
Download
Paper Nr: 202
Title:

Accelerate Training of Reinforcement Learning Agent by Utilization of Current and Previous Experience
Authors:

Chenxing Li, Yinlong Liu, Zhenshan Bing, Fabian Schreier, Jan Seyler and Shahram Eivazi
Abstract: In this paper, we examine three extensions to the Q-function Targets via Optimization (QT-Opt) algorithm and empirically studies their effects on training time over complex robotic tasks. The vanilla QT-Opt algorithm requires lots of offline data (several months with multiple robots) for training which is hard to collect in practice. To bridge the gap between basic reinforcement learning research and real world robotic applications, first we propose to use hindsight goals techniques (Hindsight Experience Replay, Hindsight Goal Generation) and Energy-Based Prioritization (EBP) to increase data efficiency in reinforcement learning. Then, an efficient offline data collection method using PD control method and dynamic buffer are proposed. Our experiments show that both data collection and training the agent for a robotic grasping task takes about one day only, besides, the learning performance maintains high level (80% successful rate). This work serves as a step towards accelerating the training of reinforcement learning for complex real world robotics tasks.
Download
Paper Nr: 203
Title:

Modeling an e-Commerce Hybrid Recommender System Based on Machine Learning Algorithms
Authors:

Antonio Panarese, Giuseppina Settanni and Angelo Galiano
Abstract: The spread of the Web and the digitalization of human society has led to the emergence of e-commerce sites. The remarkable increase in the amount of data produced by digital and automated devices forces the use of intelligent algorithms capable of processing the collected data in order to extract information. In particular, machine learning algorithms give the possibility to implement automatic models to process data and provide personalized suggestions. The advanced recommender systems are based on these models that make companies, which use the e-commerce channel, able to provide the users with suggestions on products they may be interested in. This paper proposes a model of hybrid recommender system based on the use of clustering algorithms and XGBoost, respectively, to perform a preliminary segmentation of item-customer data and predict user preference. The implemented model is discussed and preliminarily validated through a test performed using the data of a statistical sample made up of regular users of an e-commerce site.
Download
Paper Nr: 205
Title:

Logic + Reinforcement Learning + Deep Learning: A Survey
Authors:

Andreas Bueff and Vaishak Belle
Abstract: Reinforcement learning has made significant strides in recent years, including in the development of Atari and Go-playing agents. It is now widely acknowledged that logical syntax adds considerable flexibility in both the modelling of domains as well as the interpretability of domains. In this survey paper, we cover the fundamentals of how logic, reinforcement learning, and deep learning can be unified, with some ideas for future work.
Download
Paper Nr: 207
Title:

German BERT Model for Legal Named Entity Recognition
Authors:

Harshil Darji, Jelena Mitrović and Michael Granitzer
Abstract: The use of BERT, one of the most popular language models, has led to improvements in many Natural Language Processing (NLP) tasks. One such task is Named Entity Recognition (NER) i.e. automatic identification of named entities such as location, person, organization, etc. from a given text. It is also an important base step for many NLP tasks such as information extraction and argumentation mining. Even though there is much research done on NER using BERT and other popular language models, the same is not explored in detail when it comes to Legal NLP or Legal Tech. Legal NLP applies various NLP techniques such as sentence similarity or NER specifically on legal data. There are only a handful of models for NER tasks using BERT language models, however, none of these are aimed at legal documents in German. In this paper, we fine-tune a popular BERT language model trained on German data (German BERT) on a Legal Entity Recognition (LER) dataset. To make sure our model is not overfitting, we performed a stratified 10-fold cross-validation. The results we achieve by fine-tuning German BERT on the LER dataset outperform the BiLSTM-CRF+ model used by the authors of the same LER dataset. Finally, we make the model openly available via HuggingFace.
Download
Paper Nr: 213
Title:

Popularity Prediction for New and Unannounced Fashion Design Images
Authors:

Danny W. L. Yu, Eric W. T. Ngai and Maggie C. M. Lee
Abstract: People following the latest fashion trends gives importance to the popularity of fashion items. To estimate this popularity, we propose a model that comprises feature extraction using Inception v3 (a kind of Convolutional Neural Network) and a popularity score estimation using Multi-Layer Perceptron regression. The model is trained using datasets from Amazon (5,166 items) and Instagram (98,735 items) and evaluated by using mean-squared error, which is one of the many metrics of the performance of our model. Results show that, even with a simpler structure and requiring less input, our model is comparable with other more complicated methods. Our approach allows designers and manufacturers to predict the popularity of design drafts for fashion items, without exposing the unannounced design at social media or comparing with a large quantity of other items.
Download
Paper Nr: 214
Title:

Automated Deep Learning Based Answer Generation to Psychometric Questionnaire: Mimicking Personality Traits
Authors:

Anirban Lahiri, Shivam Raj, Utanko Mitra, Sunreeta Sen, Rajlakshmi Guha, Pabitra Mitra, Partha P. Chakrabarti and Anupam Basu
Abstract: Questionnaires have traditionally been used for psychometric testing and evaluation of personality traits. This work explores if personality traits or characteristics can be emulated by a computer through the responses to a questionnaire. A state-of-art Deep Learning model using natural language processing techniques coupled to a personality prediction model has been exploited. A standard OCEAN – Five-Factor evaluation questionnaire was used as the test bench for this novel study combining psychometry and machine learning. This article explains the design details of the emulation framework, the obtained results and their significance. The obtained results look promising and the framework can potentially find commercial or academic application in the near future.
Download
Paper Nr: 215
Title:

SEQUENT: Towards Traceable Quantum Machine Learning Using Sequential Quantum Enhanced Training
Authors:

Philipp Altmann, Leo Sünkel, Jonas Stein, Tobias Müller, Christoph Roch and Claudia Linnhoff-Popien
Abstract: Applying new computing paradigms like quantum computing to the field of machine learning has recently gained attention. However, as high-dimensional real-world applications are not yet feasible to be solved using purely quantum hardware, hybrid methods using both classical and quantum machine learning paradigms have been proposed. For instance, transfer learning methods have been shown to be successfully applicable to hybrid image classification tasks. Nevertheless, beneficial circuit architectures still need to be explored. Therefore, tracing the impact of the chosen circuit architecture and parameterization is crucial for the development of beneficially applicable hybrid methods. However, current methods include processes where both parts are trained concurrently, therefore not allowing for a strict separability of classical and quantum impact. Thus, those architectures might produce models that yield a superior prediction accuracy whilst employing the least possible quantum impact. To tackle this issue, we propose Sequential Quantum Enhanced Training (SEQUENT) an improved architecture and training process for the traceable application of quantum computing methods to hybrid machine learning. Furthermore, we provide formal evidence for the disadvantage of current methods and preliminary experimental results as a proof-of-concept for the applicability of SEQUENT.
Download
Paper Nr: 217
Title:

Question Difficulty Decision for Automated Interview in Foreign Language Test
Authors:

Minghao Lin, Koichiro Ito and Shigeki Matsubara
Abstract: As globalization continues to progress all over the world, demand is growing for objective and rapid assessment of language proficiency in foreign language learners. While automated assessment of listening, reading, and writing skills has been proposed, little research has been done to automate assessment of speaking skills. In this paper, we propose a method of deciding the difficulty of questions generated for interview tests of skills assessment in speaking a foreign language. To address question difficulty flexibly according to the abilities of test takers, our method considers the appropriateness of responses from test takers. We implemented this method using the large-scale pre-trained language model BERT (Bidirectional Encoder Representation from Transformers). Experiments were conducted using simulated test data from the Japanese Learner’s Conversation Database to confirm the effectiveness of our method in deciding difficulty.
Download
Paper Nr: 219
Title:

An Artificial Dendritic Neuron Model Using Radial Basis Functions
Authors:

Zachary Hutchinson
Abstract: The dendrites of biological neurons are computationally active. They contribute to the expressivity of the neural response. Thus far, dendrites have not seen wide use by the AI community. We propose a dendritic neuron model based on the compartmentalization of non-isopotential dendrites using radial basis functions. We show it is capable of producing Boolean behavior. Our goal is to grow the AI conversation around more complex neuron models.
Download
Paper Nr: 220
Title:

Supervised Learning for Untangling Braids
Authors:

Alexei Lisitsa, Mateo Salles and Alexei Vernitski
Abstract: Untangling a braid is a typical multi-step process, and reinforcement learning can be used to train an agent to untangle braids. Here we present another approach. Starting from the untangled braid, we produce a dataset of braids using breadth-first search and then apply behavioral cloning to train an agent on the output of this search. As a result, the (inverses of) steps predicted by the agent turn out to be an unexpectedly good method of untangling braids, including those braids which did not feature in the dataset.
Download
Paper Nr: 221
Title:

Multi-Graph Encoder-Decoder Model for Location-Based Character Networks in Literary Narrative
Authors:

Avi Bleiweiss
Abstract: In the past decade, an extended line of research developed a broad range of methods for reasoning about narrative from a social perspective. This often revolved around transforming literary text into a character network representation. However, there remain inconsistent traits of narrative structure produced computationally by either neural language technology or network theory tools. In this paper, we propose an encoder-decoder model with a main objective to mitigate the apparent computational divergence. Our encoder novelty lies in generating hundreds of location-based network graphs to render a fine-grained narrative. We further formalize a decoder task for detecting character communities and analyze modularity and membership affiliation. Through empirical experiments, we present visualization of stages in our computational process for four literary fiction novels.
Download
Paper Nr: 223
Title:

Improved Encoding of Possibilistic Networks in CNF Using Quine-McCluskey Algorithm
Authors:

Guillaume Petiot
Abstract: Compiling Possibilistic Networks consists in evaluating the effect of evidence by encoding the possibilistic network in the form of a multivariable function. This function can be factored and represented by a graph which allows the calculation of new conditional possibilities. Encoding the possibilistic network in Conjunctive Normal Form (CNF) makes it possible to factorize the multivariable function and generate a deterministic graph in Decomposable Negative Normal Form (d-DNNF) whose computation time is polynomial. The challenge of compiling possibilistic networks is to minimize the number of clauses and the size of the d-DNNF graph to guarantee the lowest possible computation time. Several solutions exist to reduce the number of CNF clauses. We present in this paper several improvements for the encoding of possibilistic networks. We will then focus our interest on the use of Quine-McCluskey’s algorithm (QMC) to simplify and reduce the number of clauses.
Download
Paper Nr: 226
Title:

Measures of Lexical Diversity and Detection of Alzheimer’s Using Speech
Authors:

Muskan Kothari, Darshil V. Shah, Moulya T., Swasthi P. Rao and Jayashree R.
Abstract: Alzheimer’s disease is the most common cause of dementia — a continuous decline in thinking, behavioral and social skills that affects a person’s ability to function independently. Another area of concern is the overlap of symptoms with a similar disease of dementia - Frontotemporal Dementia(FTD). This paper aims to analyze the difference in linguistic features between control and dementia groups with respect to lexical diversity through measures like Brunet’s and Sichel’s measure, frequency rates of adverb, verb, and linguistic deterioration through repetition, disfluency, incomplete sentences, hesitation and long pauses through dataset obtained by DementiaBank. This is achieved through gauging the cognitive ability in speech, which is an inexpensive and non-invasive mode of analysis, qualifying as a screening test. The subjects are given certain description tasks such as the famous cookie theft picture, analyzed through conversations. The result displays the difference in lexical diversity which is a significant marker.
Download
Paper Nr: 229
Title:

Blending Dependency Parsers With Language Models
Authors:

Nicos Isaak
Abstract: Recent advances in the AI field triggered an increasing interest in tackling various NLP tasks with language models. Experiments on several benchmarks demonstrated their effectiveness, potential, and role as a central component in modern AI systems. However, when training data are limited, specialized expertise is needed in order to help them perform complex kind-of-reasoning and achieve better results. It seems that extensive experiments with additional semantic analysis and fine-tuning are needed to achieve improvements on downstream NLP tasks. To address this complex problem of achieving better results with language models when training data are limited, we present a simplified way that automatically improves their learned representations with extra-linguistic knowledge. To this end, we show that further fine-tuning with semantics from state-of-the-art dependency parsers improves existing language models on specialized downstream tasks. Experiments on benchmark datasets we undertook show that the blending of language models with dependency parsers is promising for achieving better results.
Download
Paper Nr: 230
Title:

A Zero-Shot Transformer Model For an Attribute-Guided Challenge
Authors:

Nicos Isaak
Abstract: The Taboo Challenge competition, a task based on the well-known Taboo game, has been proposed to stimulate research in the AI field. The challenge requires building systems able to comprehend the implied inferences between the exchanged messages of guesser and describer agents. A describer sends pre-determined hints to guessers indirectly describing cities, and guessers are required to return the matching cities implied by the hints. Climbing up the scoring ledger requires resolving the highest number of cities with the smallest number of hints in a specified time frame. Here, we present TabooLM, a language-model approach that tackles the challenge based on a zero-shot setting. We start by presenting and comparing the results of our approach with three studies from the literature. The results show that TabooLM achieves SOTA results on the Taboo challenge, suggesting that it can guess the implied cities faster and more accurately than existing approaches.
Download
Paper Nr: 234
Title:

Model for Real-Time Subtitling from Spanish to Quechua Based on Cascade Speech Translation
Authors:

Abraham Alvarez-Crespo, Diego Miranda-Salazar and Willy Ugarte
Abstract: The linguistic identity of many indigenous peoples has become relevant in recent years. The speed with which many of these have been lost faces many countries of the world with a serious reduction of their cultural heritage. In South America, the critical situation of vulnerability of many of its native languages is alarming. Even languages such as Quechua, widely spoken in the region, face an early disappearance due to a low rate of inter-generational transmission. Aware of this problem, we have proposed the development of a translation and subtitling system for short film videos from Spanish to Quechua. The proposal contemplates the use of cinema to promote and retain the language. For its realization, we have built a solution that combines a Spanish Voice Recognition system and our proposal for a Quechua Machine translation model. This will be integrated with a desktop application that will also subtitle the film videos. In the tests we carry out, we have obtained better translation indicators than past proposals; in addition to the validation of Quechua-speaking users of the tool’s value. Aware of this problem, we have proposed a speech-to-text translation model that could be used as a resource for language revitalization. For its realization, we developed a cascade architecture that combines a Spanish speech recognition module and our proposal of a Quechua machine translation module, fine-tuned from a Turkish NMT model and a parallel public dataset. Additionally, we developed a subtitling algorithm to be joined with our solution into a real-time subtitling desktop application for clips of films. In the tests we carry out, we have obtained better BLEU and chrF scores than previous proposals; in addition to the validation of the translation returned in the subtitles by the Quechua speakers consulted.
Download
Paper Nr: 235
Title:

Decision-Making in Hearthstone Based on Evolutionary Algorithm
Authors:

Eiji Sakurai and Koji Hasebe
Abstract: Hearthstone is a two-player turn-based collectible card game with hidden information and randomness. Generally, the search space for actions of this game grows exponentially because the players must perform a series of actions by selecting each action from many options in each turn. When playing such a game, it is often difficult to use a game tree search technique to find the optimal sequence of actions up until the end of a turn. To solve this problem, we propose a method to determine a series of actions in Hearthstone based on an evolutionary algorithm called the rolling horizon evolutionary algorithm (RHEA). To apply RHEA to this game, we modify the genetic operators and add techniques for selecting actions based on previous search results and for filtering (pruning) some of the action options. To evaluate the effectiveness of these improvements, we implemented an agent based on the proposed method and played it against an agent based on the original RHEA for several decks. The result showed a maximum winning rate of over 97.5%. Further, our agent played against the top-performing agents in previous competitions and outperformed most of them.
Download
Paper Nr: 249
Title:

A Convolutional Neural Network Based Patch Classifier Using Mammograms
Authors:

Yumnah Hasan, Aidan Murphy, Meghana Kshirsagar and Conor Ryan
Abstract: Breast Cancer is the most prevalent cancer among females worldwide. Early detection is key to good prognosis and mammography is the most widely-used technique, particularly in screening programs. However, mammography is a highly-skilled and often time-consuming task. Deep learning methods can facilitate the detection process and assist clinicians in disease diagnosis. There has been much research showing Deep Neural Networks’ successful use in medical imaging to predict early and accurate diagnosis. This paper proposes a patch-based Convolutional Neural Network (CNN) classification approach to classify patches (small sections) obtained from mammogram images into either benign or malignant cases. A novel patch extraction approach method, which we call Overlapping Patch Extraction, is developed and compared with the two different techniques, Non-Overlapping Patch Extraction, and a Region-Based-Extraction. Experimentation is conducted using images from the Curated Breast Imaging Subset of Digital Database for Screening Mammography. Five deep learning models, three configurations of EfficientNet-V2 (B0, B2, and L), ResNet-101, and MobileNetV3L, are trained on the patches extracted using the discussed methods. Preliminary results indicate that the proposed patch extraction approach, Overlapping, produces a more robust patch dataset. Promising results are obtained using the Overlapping patch extraction technique trained on the EfficientNet-V2L model achieving an AUC of 0.90.
Download
Paper Nr: 253
Title:

Using Neural Network Architectures for Intraday Trading in the Gold Market
Authors:

Srinivas Devarajula, Vitaliy Milke and Cristina Luca
Abstract: Financial market forecasting is used to assess the future value of financial instruments in various exchange and over-the-counter markets. Investors have a high interest in the most accurate prediction of the financial instruments’ prices. Inaccurate forecasting might result in a significant financial loss in certain circumstances. This research aims to determine the most probabilistic deep learning model that can improve price forecasting in the financial markets. In this research, Convolutional Neural Networks and Long Short-Term Memory are used for the experiments to forecast the Gold price movements on the Forex market. The Gold(XAU/USD) dataset is used in this research to predict the prices for the next minute. The models proposed have been evaluated using Mean squared error, Mean absolute error, and Mean absolute percentage error metrics. The results show that the Convolutional Neural Network has performed better than the Long Short-Term Memory network and has the potential to predict the price for next minute with a low error rate.
Download
Paper Nr: 255
Title:

Software Requirements Prioritisation Using Machine Learning
Authors:

Arooj Fatima, Anthony Fernandes, David Egan and Cristina Luca
Abstract: Prioritisation of requirements for a software release can be a difficult and time-consuming task, especially when the number of requested features far outweigh the capacity of the software development team and difficult decisions have to be made. The task becomes more difficult when there are multiple software product lines supported by a software release, and yet more challenging when there are multiple business lines orthogonal to the product lines, creating a complex set of stakeholders for the release including product line managers and business line managers. This research focuses on software release planning and aims to use Machine Learning models to understand the dynamics of various parameters which affect the result of software requirements being included in a software release plan. Five Machine Learning models were implemented and their performance evaluated in terms of accuracy, F1 score and K-Fold Cross Validation (Mean).
Download
Paper Nr: 263
Title:

Creating a Math Expression Filter to Extract Concise Math Expression Images
Authors:

Kuniko Yamada and Harumi Murakami
Abstract: Even though the web is an effective resource to search for math expressions, finding appropriate ones among the obtained documents is time-consuming. Therefore, we propose a math expression filter that presents appropriate images for such searches. We call an appropriate math expression a concise math expression, such as ˆfh(x) = 1Nh ∑ni=1K(x−xih), written in a compact form whose content can be interpreted by the math expression itself. We determined the conditions satisfied by a concise math expression and developed classifiers that discriminate the images of concise math expressions from web images using supervised machine learning methods based on these conditions. We performed two experiments: Experiment 1 used methods other,than deep learning, and Experiment 2 used deep learning. A convolutional neural network (CNN) with transfer learning and fine tuning by VGG16 shows high performance with an obtained F-measure of 0.819. We applied this filter to a task that presents math expression images by entering mathematical terms into a web search engine as queries. All of the evaluation metrics outperformed the previous study, including F-measure, MAP, and MRR.
Download
Paper Nr: 271
Title:

Playstyle Generation for Geister With Genetic Algorithm and Clustering
Authors:

Keisuke Tomoda and Koji Hasebe
Abstract: Studies on game-playing agents have made various attempts to develop agents with characteristic playstyle. Most of these studies either generated agents with predetermined playstyles or simultaneously generated different playstyles without defining a specific playstyle for single-player complete information games. However, the generation of agents with different playstyles for multi-player imperfect information games has not been thoroughly investigated. Therefore, in this study, we have proposed an automatic playstyle generation method for a two-player imperfect information game called Geister. The basic idea is to use a genetic algorithm to optimize agents whose genes represent parameters that determine the manner of guessing hidden information. By clustering the genes with high fitness, obtained using this process, agents with different playstyles are generated simultaneously. From the results of the experiments, our proposed method generated five different playstyles with cyclic dominance relationships.
Download
Paper Nr: 275
Title:

Improving the Engagement of Participants by Interacting with Agents who Are Persuaded During Decision-Making
Authors:

Yoshimasa Ohmoto and Hikaru Nakaaki
Abstract: A collaborative task execution by leveraging the respective strengths of human and artificial agents can solve problems more effectively than if both entities work separately. The important point is that humans have the mental attitude to transform their own opinions through active interaction with the agent. However, people do not often interact actively with agents because they often consider them to be mere information providers. In this study, our idea was to increase participant engagement by having participants experience interactions in which the agent is persuaded by the participant in a decision-making task. We performed an experiment to analyze whether the interaction with an agent that implements a persuasion interaction model could enhance the user’s sense of self-efficacy and engagement in the task. As a result, the participant’s behavior and the questionnaire’s findings revealed that persuading the interaction partner generally improves engagement in the interaction. On the other hand, it was suggested that the experience of persuading the interaction partner and the experience of the partner agreeing influenced the subsequent engagement and subjective evaluation of the interaction.
Download
Paper Nr: 276
Title:

Improving the Human-Likeness of Game AI’s Moves by Combining Multiple Prediction Models
Authors:

Tatsuyoshi Ogawa, Chu-Hsuan Hsueh and Kokolo Ikeda
Abstract: Strong game AI’s moves are sometimes strange or difficult for humans to understand. To achieve better human-computer interaction, researchers try to create human-like game AI. For chess and Go, supervised learning with deep neural networks is one of the most effective methods to predict human moves. In this study, we first show that supervised learning is also effective in Shogi (Japanese chess) to predict human moves. We also find that the AlphaZero-based model more accurately predicted moves of players with higher skill. We then investigate two evaluation metrics for measuring human-likeness, where one is move-matching accuracy that is often used in existing works, and the other is likelihood (the geometric mean of human moves’ probabilities predicted by the model). To create game AI that is more human-like, we propose two methods to combine multiple move prediction models. One uses a Classifier to select a suitable prediction model according to different situations, and the other is Blend that mixes probabilities from different prediction models because we observe that each model is good at some situations where other models cannot predict well. We show that the Classifier method increases the move-matching accuracy by 1%-3% but fails to improve the likelihood. The Blend method increases the move-matching accuracy by 3%-4% and the likelihood by 2%-5%.
Download
Paper Nr: 279
Title:

Detecting 2D NMR Signals Using Mask RCNN
Authors:

Hadeel S. Alghamdi, Alexei Lisitsa, Igor Barsukov and Rudi Grosman
Abstract: Picking peaks in two-dimensional Nuclear Magnetic Resonance (NMR) spectra has been a critical research problem and a very time-consuming important step in further analyses of NMR biological molecular systems. Here, we implemented machine learning approach for peak detection and segmentation using machine learning framework Mask R-CNN.The model was trained on a large number of synthetic spectra of known configurations, and we show that our model demonstrates promising results up to 0.93 accuracy. We implemented uniform scaling on the data matrix during training to further improve detection to achieve 10.17% FPs and 1.7% FNs rate. We show the utility of Mask R-CNN on NMR spectra where the data range plays an important role in peak detection.
Download
Paper Nr: 286
Title:

Task Scheduling: A Reinforcement Learning Based Approach
Authors:

Ciprian Paduraru, Catalina C. Patilea and Stefan Iordache
Abstract: Nowadays, various types of digital systems such as distributed systems, cloud infrastructures, industrial devices and factories, and even public institutions need a scheduling engine capable of managing all kinds of tasks and jobs. As the global resource demand is unprecedented, we can classify task scheduling as a hot topic in today’s world. On a small scale, this process can be orchestrated by humans without the intervention of machines and algorithms. However, with large scale data streams, the scheduling process can easily exceed human capacity. An automated agent or robot capable of processing millions of requests per second is the ideal solution for efficient scheduling of flows. This work focuses on developing an agent that learns autonomously from experiences using reinforcement learning how to perform efficiently the scheduling process. Carefully designed environments are used to train the agent to have similar or better planning experiences than already existing methods such as heuristic algorithms, machine learning-based methods (supervised algorithms) and genetic algorithms. We also focused on designing a suitable dataset generator for the research community, a tool that generates random data starting from a user-supplied template in combination with different distribution strategies.
Download
Paper Nr: 287
Title:

pdi-Bagging: A Proposal of Bagging-Type Ensemble Method Generating Virtual Data
Authors:

Honoka Irie and Isao Hayashi
Abstract: For pattern classification problems, there is ensemble learning method that identifies multiple weak classifiers by the learning data and combines them together to improve the discrimination rate of testing data. We have already proposed pdi-Bagging (Possibilistic Data Interpolation-Bagging) which improves the discrimination rate of testing data by adding virtually generated data to learning data. In this paper, we propose a new method to specify the generation area of virtual data and change the generation class of virtual data. As a result, the discriminant accuracy is improved since five new bagging methods which generate virtual data around correct discrimination data and error discrimination data are formulated, and the class of virtual data is determined with the proposed new evaluation index in multidimensional space. We formulate a new pdi-Bagging algorithm, and discuss the usefulness of the proposed method using numerical examples.
Download
Paper Nr: 289
Title:

Hard Spatial Attention Framework for Driver Action Recognition at Nighttime
Authors:

Karam Abdullah, Imen Jegham, Mohamed A. Mahjoub and Anouar Ben Khalifa
Abstract: Driver monitoring has become a key challenge in both computer vision and intelligent transportation system research fields due to its high potential to save pedestrians, drivers, and passengers’ lives. In fact, a variety of issues related to driver action classification in real-world driving settings are present and make classification a challenging task. Recently, driver in-vehicle action relying on deep neural networks has made significant progress. Though promising classification results have been achieved in the daytime, the performance in the nighttime remains far from satisfactory. In addition, deep learning techniques treat the whole input data with the same importance which is confusing. In this work, a nighttime driver action classification network called hard spatial attention is proposed. Our approach effectively captures the relevant dynamic spatial information of the cluttered driving scenes under low illumination for an efficient driver action classification. Experiments are performed on the unique public realistic driver action dataset recorded at nighttime 3MDAD dataset. Our approach outperforms state-of-the-art methods’ classification accuracies on both side and front views.
Download
Paper Nr: 293
Title:

Turn-Based Multi-Agent Reinforcement Learning Model Checking
Authors:

Dennis Gross
Abstract: In this paper, we propose a novel approach for verifying the compliance of turn-based multi-agent reinforcement learning (TMARL) agents with complex requirements in stochastic multiplayer games. Our method overcomes the limitations of existing verification approaches, which are inadequate for dealing with TMARL agents and not scalable to large games with multiple agents. Our approach relies on tight integration of TMARL and a verification technique referred to as model checking. We demonstrate the effectiveness and scalability of our technique through experiments in different types of environments. Our experiments show that our method is suited to verify TMARL agents and scales better than naive monolithic model checking.
Download
Paper Nr: 296
Title:

Emotion-Cause Pair Extraction as Question Answering
Authors:

Huu-Hiep Nguyen and Minh-Tien Nguyen
Abstract: The task of Emotion-Cause Pair Extraction (ECPE) aims to extract all potential emotion-cause pairs of a document without any annotation of emotion or cause clauses. Previous approaches on ECPE have tried to improve conventional two-step processing schemes by using complex architectures for modeling emotion-cause interaction. In this paper, we cast the ECPE task to the question answering (QA) problem and propose simple yet effective BERT-based solutions to tackle it. Given a document, our Guided-QA model first predicts the best emotion clause using a fixed question. Then the predicted emotion is used as a question to predict the most potential cause for the emotion. We evaluate our model on a standard ECPE corpus. The experimental results show that despite its simplicity, our Guided-QA achieves promising results and is easy to reproduce. The code of Guided-QA is also provided.
Download
Paper Nr: 298
Title:

Denial of Service Detection for IoT Networks Using Machine Learning
Authors:

Husain Abdulla, Hamed S. Al-Raweshidy and Wasan Awad
Abstract: The Internet of Things (IoT) is considered one of the trending technologies today. IoT affects a variety of industries, including logistics tracking, healthcare, automotive and smart cities. A rising number of cyberattacks and breaches are rapidly targeting networks equipped with IoT devices. Due to the resource-constrained nature of the IoT devices, one of the Internet security issues impacting IoT devices is the Denial-of-Service (DoS). This encourages the development of new techniques for automatically detecting DoS in IoT networks. In this paper, we test the performance of the following Machine Learning (ML) algorithms in detecting IoT DoS attacks using packet analysis at regular time intervals: Neural Networks (NN), Gaussian Naive Bayes (NB), Decision Trees (DT), and Support Vector Machine (SVM). We were able to achieve 98% accuracy in intrusion detection for IoT devices. We have created a novel way of detecting the attacks using only six attributes, which significantly reduces the time to train the ML Models by 58% on average. This research is based on data collected from actual IoT attacks on IoT networks. This paper shows that using the DT or NN; we can detect attacks on IoT devices. Furthermore, it shows that NB and SVM are poor in detecting IoT attacks. In addition, it proves that middle boxes embedded with ML Models can be utilized to detect attacks in places such as houses, manufactures, and plants.
Download
Paper Nr: 301
Title:

Learning Preferences in Lexicographic Choice Logic
Authors:

Karima Sedki, Nada Boudegzdame, Jean B. Lamy and Rosy Tsopra
Abstract: Lexicographic Choice Logic (LCL) is a variant of Qualitative Choice Logic which is a logic-based formalism for preference handling. The LCL logic extends the propositional logic with a new connective (⃗⋄ ) to express preferences. Given a preference x⃗⋄ y, satisfying both x and y is the best option, the second best option is to satisfy only x, and satisfying only y is the third best option. Satisfying neither x nor y is not acceptable. In this paper, we propose a method for learning preferences in the context of LCL. The method is based on an adaptation of association rules based on the APRIORI algorithm. The adaptation consists essentially of using variations of the support and confidence measures that are suitable for LCL semantic.
Download
Paper Nr: 302
Title:

Sentence Transformers and DistilBERT for Arabic Word Sense Induction
Authors:

Rakia Saidi and Fethi Jarray
Abstract: Word sense induction (WSI) is a fundamental task in natural language processing (NLP) that consists in discovering the sense associated to each instance of a given target ambiguous word. In this paper, we propose a two-stage approach for solving Arabic WSI. In the first stage, we encode the input sentence into context representations using Transformer-based encoder such as BERT or DistilBERT. In the second stage, we apply clustering to the embedded corpus obtained in the first stage by using K-Means and Agglomerative Hierarchical Clustering (HAC). We evaluate our proposed method on the Arabic WSI summarization task. Experimental results show that our model achieves new state-of-the-art on both the Open Source Arabic Corpus (OSAC)(Saad and Ashour, 2010) and the SemEval arabic (2017).
Download
Paper Nr: 305
Title:

A Knowledge-Based Approach for Evaluating Impact of Therapeutic Strategies
Authors:

Nadia Abchiche-Mimouni, Mike T. Nzali and François Gueyffier
Abstract: This paper proposes an original approach for modelling medical expertise and simulating medical strategies. A knowledge-based system is used to model therapeutic strategies according to three axes: diagnostic, prescription and treatment effect. The diagnostic axis describes the ways of deciding whether an individual is eligible for treatment or not. The prescription axis models the ways of choosing an adequate drug for an individual or changing the current treatment if it is judged ineffective. Treatment effect concerns the effect of a drug at the individual level. This modelling is used for exploring different therapeutic strategies and quantifying their impact on the individual and population levels. We have developed a platform, based on a rule-based system, that was validated with a Use-case in Hypertension management. Classical and Alternative strategies have been simulated with the same Realistic virtual population. 20.000 individuals were considered and several parameters (e.g. optimal drug prescription, evolution of the cardiovascular risk) were calculated. The experiments showed the viability and relevance of the approach. Its strengths are numerous. Since the rules are the input of the system, they can be introduced and modified by non-programmers people, allowing prescribers to fully test their own rules. The platform is configurable in terms of modelled expertise and in terms of outputs to be measured. Empirical results concerning the superiority of the Alternative strategies have been produced.
Download
Paper Nr: 307
Title:

Comparative Analysis of Process Models for Data Science Projects
Authors:

Damian Kutzias, Claudia Dukino, Falko Kötter and Holger Kett
Abstract: When adopting data science technology into practice, enterprises need proper tools and process models. Data science process models guide the project management by providing workflows, dependencies, requirements, relevant challenges and questions as well as suggestions of proper tools for all tasks. Whereas process models for classic software development have evolved for a comparably long time and therefore have a high maturity, data science process models are still in rapid evolution. This paper compares existing data science process models using literature analysis, and identifies the gap between existing models and relevant challenges by performing interviews with experts.
Download
Paper Nr: 309
Title:

Coevolving Hexapod Legs to Generate Tripod Gaits
Authors:

Cameron L. Angliss and Gary B. Parker
Abstract: This research is the first step in new research that is investigating the use of an extensive neural network (NN) system to control the gait of a hexapod robot by sending specific position signals to each of the leg actuators. The system is highly distributed with nerve clusters that control each leg and no central controller. The intent is to create a more biologically-inspired system and one that has the potential to dynamically change its gait in real-time to accommodate for unforeseen malfunctions. We approach the evolution of this extensive NN system in a unique way by treating each of the legs as an individual agent and using cooperative coevolution to evolve the team of heterogeneous leg agents to perform the task, which in this initial phase is walking forward on a flat surface. When tested using a simulated hexapod robot modeled after an actual robot, this new method reliably produced a stable tripod gait.
Download
Paper Nr: 15
Title:

Chop-SAT: A New Approach to Solving SAT and Probabilistic SAT for Agent Knowledge Bases
Authors:

Thomas C. Henderson, David Sacharny, Amar Mitiche, Xiuyi Fan, Amelia Lessen, Ishaan Rajan and Tessa Nishida
Abstract: An early approach to solve the SAT problem was to convert the disjunctions directly to equations which create an integer programming problem with 0-1 solutions. We have independently developed a similar method which we call Chop-SAT based on geometric considerations. Our position is that Chop-SAT provides a wide range of geometric approaches to find SAT and probabilistic SAT (PSAT) solutions. E.g., one potentially powerful approach to determine that a SAT solution exists is to fit the maximal volume ellipsoid and explore it semi-major axis direction to find an Hn vertex in that direction.
Download
Paper Nr: 18
Title:

Nature-Inspired Algorithms for Solving Weighted Constraint Satisfaction Problems
Authors:

Mehdi Bidar and Malek Mouhoub
Abstract: Several applications such as timetabling, scheduling and resource allocation, can be represented as a Constraint Satisfaction Problem (CSP). Solving a CSP consists in finding a complete assignment of values to variables satisfying all the constraints. In many real-life scenarios (including over-constrained problems), some constraints (called soft constraints) can be violated according to some penalty function. In this regard, the Weighted CSP (WCSP) can be used as an extension of the CSP where each constraint comes with a cost function. Solving a WCSP consists in finding an optimal solution minimizing the total costs related to all constraints. Searching for an optimal solution to a WCSP is usually dealt with classical complete methods like backtracking and bucket elimination techniques. However, since WCSPs are NP-hard, complete methods will require exponential time cost. Therefore, approximation methods such as metaheuristics are appropriate alternatives as they are capable of providing a good compromise between the quality of the solution and the corresponding running time. We study the applicability of several nature-inspired techniques including; Particle Swarm Optimization (PSO), Firefly, Genetic Algorithms (GAs), Artificial Bee Colony (ABC), Mushroom Reproduction Optimization (MRO), Harmony Search (HS) and Focus Group (FG). While these methods do not guarantee the optimality of the solution returned, they are in general successful in returning a good solution in a desirable time cost. This statement has been demonstrated through the experimental results we conducted on randomly generated WCSP instances following the known RB model. The latter has been adopted as it has the ability to produce hard-to-solve random problem instances. The obtained results are promising and show the potential of the considered nature-inspired techniques.
Download
Paper Nr: 21
Title:

Explainable Recommendations of Drugs for Diabetic Patients
Authors:

Priscila Valdiviezo-Diaz
Abstract: Currently, recommender systems are widely used for different purposes, for example, to recommend resources, products, and services. In the health domain, recommender systems are being used to recommender drugs, treatments, food plans, and healthcare services in general. Collaborative filtering is the most popular technique in the recommender system area. This technique can be of two types: memory-based collaborative filtering and based-model collaborative filtering. One of the problems of recommender systems is that most of them focus on enhancing the precision of the recommendation and do not provide a justification for the suggestions given to the user. Therefore, it is important to provide explainable recommendations so that the user understands why an item is recommended. To address this problem, in this paper the use of a Bayesian method for explainable drug recommendations for diabetic patients is presented. Several experiments are carried out using a dataset with information on diabetic patients with three collaborative filtering approaches: the memory-based approach IbCF, and two model-based approaches: item-based NBCF, and Hybrid NBCF. The experimental results present good results for the Hybrid NBCF approach compared to the other approaches tested. Moreover, it is observed a better quality of prediction and an increase in recommendation precision with Hybrid NBCF.
Download
Paper Nr: 24
Title:

Automatic Reactive Power Classification Based on Selected Machine Learning Methods
Authors:

Viktor Pristaš, L’ubomír Antoni and Gabriel Semanišin
Abstract: Reactive power is an important part of the electric power systems in order to rotate machines or to transmit active power by transmission lines. However, an excess of reactive power in the electrical systems can increase the risk of failure of the transmission system. We present an automatic reactive power classification on multifamily residential dataset of electricity based on selected machine learning methods. We aim to predict an excess of reactive power in the apartments located in the Northeastern United States. Moreover, we explore the statistical significance of differences between mean performances of selected machine learning methods.
Download
Paper Nr: 28
Title:

A Sequence-to-Sequence Neural Network for Joint Aspect Term Extraction and Aspect Term Sentiment Classification Tasks
Authors:

Hasna Chouikhi, Fethi Jarray and Mohammed Alsuhaibani
Abstract: Aspect-based Sentiment Analysis (ABSA) consists in extracting the terms or entities described in a text (attributes of a product or service) and the user perception of each aspect. Most earlier approaches are traditionally programmed sequentially, extracting the terms and then predicting their polarity. In this paper, we propose a joint sequence-to-sequence model that simultaneously extracts the terms and determines their polarities. The seq2seq architecture comprises an encoder, which can be an Arabic BERT model, and a decoder, which also can be an Arabic BERT, GPT, or BiGRU model. The encoder aims to preprocess the input sequence and encode it into a fixed-length vector called a context vector. The decoder reads that context vector from the encoder and generates the aspect term sentiment pair output sequence. We conducted experiments on two accessible Arabic datasets: Human Annotated Arabic Dataset (HAAD) of Book Reviews and The ABSA Arabic Hotels Reviews (ABSA Arabic Hotels). We achieve an accuracy score of 77% and 96% for HAAD and ABSA Arabic Hotels datasets respectively using BERT2BERT pairing. The results clearly highlight the superiority of the joint seq2seq model over pipeline approaches and the outperformance of BERT2BERT architecture over the pairing of BERT and BiGRU, and the pairing of BERT and GPT.
Download
Paper Nr: 33
Title:

ECA-CE: An Evolutionary Clustering Algorithm with Initial Population by Clustering Ensemble
Authors:

Chao Xu, Chunlin Xu and Shengli Wu
Abstract: Evolutionary clustering is a type of algorithm that uses genetic algorithms to optimize clustering results. Unlike traditional clustering algorithms which obtain clustering results by iteratively increasing the distance between clusters and reducing the distance between instances within a cluster, the evolutionary clustering algorithm tries to search for the optimal clustering result in the solution space. Not surprisingly, the initial population set in an evolutionary clustering algorithm has significant influence on the final results. To ensure the quality of the initial population, this paper proposed a clustering ensemble-based method, ECA-CE, to do the initial population for the evolutionary clustering algorithm. In ECA-CE, a clustering ensemble method, Hybrid Bipartite Graph Formulation, is applied. Extensive experiments are conducted on 20 benchmark datasets, and the experimental results demonstrate that the proposed ECA-CE is more effective than two evolutionary clustering algorithms F1-ECAC and ECAC in terms of Adjusted Rand index.
Download
Paper Nr: 55
Title:

MAKI: A Multi-Agent Public Key Infrastructure
Authors:

Arthur Baudet, Oum-El-kheir Aktouf, Annabelle Mercier and Philippe Elbaz-Vincent
Abstract: This paper presents a public key infrastructure for open multi-agent systems of embedded agents. These systems rely on agents autonomy and cooperation between heterogeneous systems to achieve common goals. They are very prone to attacks as they are confronted with unknown systems with unknown goals. We aim at securing the communications between agents to provide foundations for more advanced security solutions as well as allowing agents to communicate without the risk of their messages being tampered with. To do so, we deploy a key infrastructure taking advantage of the agents autonomy to allow authenticity and integrity checks and accountability of all interactions. The result of our approach is the Multi-Agent Key Infrastructure, a public key infrastructure leveraging and empowering a trust management system. Agents autonomously maintain a set of trusted certification authorities that deliver certificates to trusted agents and revoke intruders. This infrastructure paves the way to build more secure open decentralized systems of autonomous embedded systems. To make our solution general and adaptable to many situations, we only provide recommendations related to the cryptographic primitives to use and the inner details of the trust management system.
Download
Paper Nr: 57
Title:

On the Prediction of a Nonstationary Exponential Distribution Based on Bayes Decision Theory
Authors:

Daiki Koizumi
Abstract: A prediction problem with a nonstationary exponential distribution based on the Bayes decision theory was considered in this paper. The proposed predictive algorithm is based on both posterior and predictive distributions in a Bayesian context. The predictive estimator satisfies the Bayes optimality, which guarantees a minimum average error rate with a nonstationary probability model, a squared loss function, and a prior distribution of parameter. Finally, the predictive performance of the proposed algorithm was evaluated via comparison with the stationary exponential distribution using real meteorological data.
Download
Paper Nr: 66
Title:

Constraint-Based Filtering and Evaluation of CSP Search Trees
Authors:

Maximilian Bels, Sven Löffler, Ilja Becker and Petra Hofstedt
Abstract: Using Constraint Programming (CP) real world problems can be described conveniently in a declarative way with constraints in a so-called constraint satisfaction problem (CSP). Finite domain CSPs (FD-CSPs) are one form of CSPs, where the domains of the variables are finite. Such FD-CSPs are mostly evaluated by a search nested with propagation, where the search process can be represented by search trees. Since search can quickly become very time-consuming, especially with large variable domains (solving CSPs is NP-hard in general), heuristics are used to control the search, which in many cases — depending on the problem — allow to achieve a performance gain. In this paper, we present a new method for filtering and evaluating search trees of FD-CSPs. Our new tree filtering method is based on the idea of formulating and evaluating filters as constraints over FD-CSP search trees. The constraint-based formulation of filter criteria proves to be very flexible. Our new technique was integrated into the Visual Constraint Solver (VCS) tool, which allows the solution process of CSPs to be followed interactively and step by step through a suitable visualization.
Download
Paper Nr: 68
Title:

Towards Automatic Medical Report Classification in Czech
Authors:

Pavel Přibáň, Josef Baloun, Jiří Martínek, Ladislav Lenc, Martin Prantl and Pavel Král
Abstract: This paper deals with the automatic classification of medical reports in the form of unstructured texts in Czech. The outcomes of this work are intended to be integrated into a coding assistant, a system that will help the clinical coders with the manual coding of the diagnoses. To solve this task, we compare several approaches based on deep neural networks. We compare the models in two different scenarios to show their advantages and drawbacks. The results demonstrate that hierarchical GRU with attention outperforms all other models in both cases. The experiments further show that the system can significantly reduce the workload of the operators and thus also saves time and money. To the best of our knowledge, this is the first attempt at automatic medical report classification in the Czech language.
Download
Paper Nr: 72
Title:

A Sequence-Motif Based Approach to Protein Function Prediction via Deep-CNN Architecture
Authors:

Vikash Kumar, Ashish Ranjan, Deng Cao, Gopalakrishnan Krishnasamy and Akshay Deepak
Abstract: The challenge of determining protein functions, inferred from the study of protein sub-sequences, is a complex problem. Also, a little literature is evident in this regard, while a broad coverage of the literature shows a bias in the existing approaches for the full-length protein sequences. In this paper, a CNN-based architecture is introduced to detect motif information from the sub-sequence and predict its function. Later, functional inference for sub-sequences is used to facilitate the functional annotation of the full-length protein sequence. The results for the proposed approach demonstrate a great future ahead for further exploration of sub-sequence based protein studies. Comparisons with the ProtVecGen-Plus – a (multi-segment + LSTM) approach – demonstrate, an improvement of +1.24% and +4.66% for the biological process (BP) and molecular function (MF) subontologies, respectively. Next, the proposed method outperformed the hybrid ProtVecGen-Plus + MLDA by a margin of +3.45% for the MF dataset, while raked second for the BP dataset. Overall, the proposed method produced better results for significantly large protein sequences (having sequence length > 500 amino acids).
Download
Paper Nr: 83
Title:

Continual Optimization of In-Production Machine Learning Systems Through Semantic Analysis of User Feedback
Authors:

Hemadri Jayalath, Ghadeer Yassin, Lakshmish Ramaswamy and Sheng Li
Abstract: With the rapid advancement of machine learning technologies, a wide range of industries and domains have adopted machine learning in their key business processes. Because of this, it is critical to ensure the optimal performance of operationalized machine learning models. This requires machine learning models to be regularly monitored and well-maintained after deployment. In this paper, we discuss the benefits of getting human guidance during the machine learning model maintenance stage. We present a novel approach that semantically evaluates end-user feedback and identifies the sentiment of the users towards different aspects of machine learning models and provides guidance to systematize the fixes according to the priority. We also crawled the web and created a small data set containing user feedback related to language models and evaluated it using our approach and uncovered some interesting insights related to language models. Further, we compare the trade-offs of alternative techniques that can be applied in different stages in our proposed model evaluation pipeline. Finally, we have provided insights and the future work that can be done to broaden the area of machine learning maintenance with human collaboration.
Download
Paper Nr: 85
Title:

Multi-Objective Optimization of the Dynamic and Flexible Job Shop Scheduling Problem Under Workers Fatigue Constraints
Authors:

Dorsaf Aribi, Olfa Belkahla Driss and Hind B. El Haouzi
Abstract: A massive number of studies has tackled the scheduling problem, but they mainly seek to solve the classic problem by reducing the real constraints of the environment like workers’ fatigue, which may lead to defective production, and the occurrence of unexpected events that makes the initial scheduling obsolete. In this paper, we study the multi-objective dynamic flexible job shop-scheduling problem under workers’ fatigue constraints (DFJSP-WF) through three unexpected events: job insertion, machine breakdown and job cancellation. First, a multi-objective model is established with objectives to minimize makespan and total weighted tardiness, earliness and rejected parts due to workers’ errors, which depend on workers’ fatigue. Second, to deal with this model, a non-dominated sorting genetic algorithm II (NSGA-II) is adapted. Computational results are presented using three sets of well-known benchmark literature instances.
Download
Paper Nr: 87
Title:

A New Approach to Moving Object Detection and Segmentation: The XY-shift Frame Differencing
Authors:

N. Wondimu, U. Visser and C. Buche
Abstract: Motion out-weights other low-level saliency features in attracting human attention and defining region of interests. The ability to effectively identify moving objects in a sequence of frames help to solve important computer vision problems, such as moving object detection and segmentation. In this paper, we propose a novel frame differencing technique along with a simple three-stream encoder-decoder architecture to effectively and efficiently detect and segment moving objects in a sequence of frames. Our frame differencing component incorporates a novel self-differencing technique, which we call XY-shift frame differencing, and an improved three-frame differencing technique. We fuse the feature maps from the raw frame and the two outputs of our frame differencing component, and fed them to our transfer-learning based convolutional base, VGG-16. The result from this sub-component is further deconvolved and the desired segmentation map is produced. The effectiveness of our model is evaluated using the re-labeled multi-spectral CDNet-2014 dataset for motion segmentation. The qualitative and quantitative results show that our technique achieves effective and efficient moving object detection and segmentation results relative to the state-of-the-art methods.
Download
Paper Nr: 90
Title:

Proposal of a Signal Control Method Using Deep Reinforcement Learning with Pedestrian Traffic Flow
Authors:

Akimasa Murata, Yuichi Sei, Yasuyuki Tahara and Akihiko Ohsuga
Abstract: In dealing with traffic control problems, there have been studies on learning signal change patterns and timing by using reinforcement learning for signals. In most of them, the focus is on improving the delay time of vehicles, and few of them assume the traffic situation including pedestrians. Therefore, the objective of this study is to provide traffic control to reduce traffic delays for both vehicles and pedestrians in an environment where pedestrian traffic volume varies greatly. Then, we will verify the accuracy with traffic signals considering the temporal changes of the environment. Results of verification, although vehicle wait times increased, a significant reduction in pedestrian wait times was observed.
Download
Paper Nr: 92
Title:

The MindSpaces Knowledge Graph: Applied Logic and Semantics on Indoor and Urban Adaptive Design
Authors:

Evangelos A. Stathopoulos, Alexandros Vassiliades, Sotiris Diplaris, Stefanos Vrochidis, Nick Bassiliades and Ioannis Kompatsiaris
Abstract: The evolution of Knowledge Graphs (KGs), during the last two decades, has encouraged developers to create more and more context related KGs. This advance is extremely important because Artificial Intelligence (AI) applications can access open domain specific information in a semantically rich, machine understandable format. In this paper, we present the MindSpaces KG, a KG that can represent emotions-relevant and functional design for the indoor and urban adaptive design. The MindSpaces KG can integrate emotional, physiological, visual, and textual measurements, for the development of online adapting environments. Moreover, we present a reasoning mechanism that extracts crucial knowledge from the MindSpaces KG, which can help users in real-life scenarios. The scenarios were provided by experts.
Download
Paper Nr: 97
Title:

VoronoiPatches: Evaluating a New Data Augmentation Method
Authors:

Steffen Illium, Gretchen Griffin, Michael Kölle, Maximilian Zorn, Jonas Nüßlein and Claudia Linnhoff-Popien
Abstract: Overfitting is a problem in Convolutional Neural Networks (CNN) that causes poor generalization of models on unseen data. To remediate this problem, many new and diverse data augmentation (DA) methods have been proposed to supplement or generate more training data, and thereby increase its quality. In this work, we propose a new DA algorithm: VoronoiPatches (VP). We primarily utilize non-linear re-combination of information within an image, fragmenting and occluding small information patches. Unlike other DA methods, VP uses small convex polygon-shaped patches in a random layout to transport information around within an image. In our experiments, VP outperformed current DA methods regarding model variance and overfitting tendencies. We demonstrate DA utilizing non-linear re-combination of information within images, and non-orthogonal shapes and structures improves CNN model robustness on unseen data.
Download
Paper Nr: 105
Title:

A Framework for Explaining Accident Scenarios for Self-Driving Cars
Authors:

Omar Wahbi, Yasmin Mansy, Nourhan Ehab and Amr Elmougy
Abstract: In the current state of the art, intelligent decision-making in autonomous vehicles is not typically comprehensible by humans. This deficiency prevents this technology from becoming socially acceptable. In fact, one of the most critical challenges that autonomous vehicles face is the need for making instantaneous decisions as there are reports of self-driving cars unnecessarily hesitating and deviating when objects are detected near the vehicle, hence possibly having car accidents. As a result, gaining a thorough understanding of autonomous vehicle reported accidents is becoming increasingly important. In addition to making real-time decisions, the autonomous car AI system must be able to explain how its decisions are made. Therefore, in this paper, we propose an explanation framework capable of providing the reasons why an autonomous vehicle made a particular decision, specifically in the occurrence of a car accident. Overall, results showed that the framework generates correct explanations for the decisions that were taken by an autonomous car by getting the nearest possible and feasible counterfactual.
Download
Paper Nr: 114
Title:

Rumor Detection in Tweets Using Graph Convolutional Networks
Authors:

Takumi Takei, Yuichi Sei, Yasuyuki Tahara and Akihiko Ohsuga
Abstract: The recent development of social networking services has made it easier for anyone to get information. On the other hand, rumors which are information whose truth is unverified are not only easy to spread but also can cause damage such as flames, incitement, and slander. Accurate identification of rumors is effective against such problems and may prevent the spread of misinformation. Based on previous research, this study created a dataset of rumors including replies to fact-checked Japanese tweets. Using a GCN-based deep learning classifier, we performed binary classification of whether a tweet is a False rumor or not, and multinomial classification of True rumor, False rumor, and Unclear rumor, varying the amount of propagation information used. The result of binary classification shows that the maximum accuracy is 0.637, and the maximum F value is 0.641, while the result of multinomial classification shows that the maximum accuracy is 0.547, and the maximum F value is 0.460. We discussed the effectiveness of propagation information and deep learning for detecting Japanese rumors.
Download
Paper Nr: 115
Title:

Potentials of Explainable Predictions of Order Picking Times in Industrial Production
Authors:

Kaja Balzereit, Nehal Soni and Andreas Bunte
Abstract: The order picking process in a manufacturing supermarket is central in many industrial productions as it ensures that the items required for production are provided at the right time. However, the order picking process itself often is a black box, i.e., the time it takes to pick an order and the dependencies in the process that influence the time usually are not exactly known. In this work, we highlight the potentials of creating explainable predictions of order picking times using Artificial Intelligence methods. The prediction is based on the analysis of a historic database and on a linear regression analysis that learns the dependencies in the data. From this prediction, (1) the potential of identifying features having a high and a low influence on the order picking time, (2) the potential of optimizing the order picking process itself, and (3) the potential of optimizing depending processes are identified. For prediction, we utilize the regression methods LASSO and Decision Tree. These methods are compared with regard to their interpretability and usability in industrial manufacturing.
Download
Paper Nr: 124
Title:

A Review of AutoML Software Tools for Time Series Forecasting and Anomaly Detection
Authors:

Christian O’Leary, Farshad G. Toosi and Conor Lynch
Abstract: Time series exist across a plethora of domains such as sensors, market prices, network traffic, and health monitoring. Modelling time series data allows researchers to perform trend analysis, forecasting, anomaly detection, predictive maintenance, and data exploration. Given the theoretical and technical knowledge required to implement mathematical and machine learning models, numerous software libraries have emerged to facilitate the programming of these algorithms via automated machine learning (AutoML). Comparatively few studies compare such technologies in the context of time series analysis and existing tools are often limited in functionality. This review paper presents an overview of AutoML software for time series data for both forecasting and anomaly detection. The analysis considers 28 metrics that indicate functionality coverage, code maturity, and community support across 22 AutoML libraries. These aspects of software development are crucial for the uptake and utilisation of AutoML tools. This study proposes a means of deriving a functionality score for correlation analysis between variables such as lines of code, package downloads from PyPi, and GitHub issue completion rate. This review paper also presents an overview of AutoML library features which can facilitate informed decisions on which tools are most appropriate in various instances.
Download
Paper Nr: 127
Title:

Catch Me if You Can: Improving Adversaries in Cyber-Security with Q-Learning Algorithms
Authors:

Arti Bandhana, Ondřej Lukáš, Sebastian Garcia and Tomáš Kroupa
Abstract: The ongoing rise in cyberattacks and the lack of skilled professionals in the cybersecurity domain to combat these attacks show the need for automated tools capable of detecting an attack with good performance. Attackers disguise their actions and launch attacks that consist of multiple actions, which are difficult to detect. Therefore, improving defensive tools requires their calibration against a well-trained attacker. In this work, we propose a model of an attacking agent and environment and evaluate its performance using basic Q-Learning, Naive Q-learning, and DoubleQ-Learning, all of which are variants of Q-Learning. The attacking agent is trained with the goal of exfiltrating data whereby all the hosts in the network have a non-zero detection probability. Results show that the DoubleQ-Learning agent has the best overall performance rate by successfully achieving the goal in 70% of the interactions.
Download
Paper Nr: 133
Title:

Randout-KD: Finetuning Foundation Models for Text Classification via Random Noise and Knowledge Distillation
Authors:

Pervaiz I. Khan, Andreas Dengel and Sheraz Ahmed
Abstract: Finetuning foundation models effectively on downstream tasks is ongoing research. In this paper, we present a finetuning method “Randout-KD” that enhances the performance of a student model for text classification. We specifically propose a noise-injecting method in the representations of the transformer model during its finetuning that works as regularization. Moreover, we integrate the knowledge distillation and noise injection methods and show that combining these approaches boosts the baseline model performance. We evaluate the proposed method on two datasets namely “CODA-19” and “RHMD” using PubMedBERT and RoBERTa Large as teacher models, and data2vec as a student model. Results show that the proposed approach improves the accuracy up to 1.2% compared to the baseline methods.
Download
Paper Nr: 134
Title:

Automated Planning for Military Airline Controller Training Scenarios
Authors:

Romain Goutiere, Domitile Lourdeaux and Sylvain Lagrue
Abstract: In this paper we focus on the generation of scenarios for military airline controller training in a virtual environment. We are using a planning system mixing temporal planning and hierarchical task networks based on the ANML planning language and allowing a better representation of both narrative and pedagogical objectives. We also propose and test a method to automatically generate potential alternative plans to the initially planned scenario, reaching the same objectives, to make it more robust.
Download
Paper Nr: 160
Title:

Multi Platform-Based Hate Speech Detection
Authors:

Shane Cooke, Damien Graux and Soumyabrata Dev
Abstract: A major issue faced by social media platforms today is the detection, and handling of hateful speech. The intricacies and imperfections of online communication make this a difficult task, and the rapidly changing use of both non-hateful, and hateful language in the online sphere means that researchers must constantly update and modify their hate speech detection methodologies. In this study, we propose an accurate and versatile multi-platform model for the detection of hate speech, using first-hand data scraped from some of the most popular social media platforms, that we share to the community. We explore and optimise 50 different model approaches, and evaluate their performances using several evaluation metrics. Overall, we successfully build a hate speech detection model, pairing the USE word embeddings with the SVC machine learning classifier, to obtain an average accuracy of 95.65% and achieved a maximum accuracy of 96.89%. We also develop and share an application allowing users to test sentences against a collection of the most accurate hate speech detection models. Our application then returns a aggregated hate speech classification, together with a confidence level, and a breakdown of the methodologies used to produce the final classification for explainability.
Download
Paper Nr: 182
Title:

Convolutional Networks Versus Transformers: A Comparison in Prostate Segmentation
Authors:

Fernando Vásconez, Maria Baldeon Calisto, Daniel Riofrío, Zhouping Wei and Yoga Balagurunathan
Abstract: Prostate cancer is one of the most common types of cancer that affects men. One way to diagnose and treat it is by manually segmenting the prostate region and analyzing its size or consistency in MRI scans. However, this process requires an experienced radiologist, is time-consuming, and prone to human error. Convolutional Neural Networks (CNNs) have been successful at automating the segmentation of the prostate. In particular, the U-Net architecture has become the de-facto standard given its performance and efficacy. However, CNNs are unable to model long-range dependencies. Transformer networks have emerged as an alternative, obtaining better results than CNNs in image analysis when a large dataset is available for training. In this work, the residual U-Net and the transformer UNETR are compared in the task of prostate segmentation on the ProstateX dataset in terms of segmentation accuracy and computational complexity. Furthermore, to analyze the impact of the size of the dataset, four training datasets are formed with 30, 60, 90, and 120 images. The experiments show that the CNN architecture has a statistical higher performance when the dataset has 90 or 120 images. When the dataset has 60 images, both architectures have a statistical similar performance, while when the dataset has 30 images UNETR performs marginally better. Considering the complexity, the UNETR has 5x more parameters and takes 5.8x more FLOPS than the residual U-Net. Therefore, showing that in the case of prostate segmentation CNNs have an overall better performance than Transformer networks.
Download
Paper Nr: 186
Title:

PARL: A Dialog System Framework with Prompts as Actions for Reinforcement Learning
Authors:

Tao Xiang, Yangzhe Li, Monika Wintergerst, Ana Pecini, Dominika Młynarczyk and Georg Groh
Abstract: The performance of most current open-domain dialog systems is limited by the (training) dialog corpora due to either generation-based or retrieval-based learning patterns. To circumvent this limitation, we propose PARL, an open-domain dialog system framework using Prompts as Actions for Reinforcement Learning. This framework requires a (fixed) open-domain dialog system as the backbone and trains a behavior policy using reinforcement learning to guide the backbone system to respond appropriately with respect to a given conversation. The action space is defined as a finite set of behaviors in the form of natural language prompts. Preliminary results show that with the guidance of the behavior policy, the backbone system could generate more engaging and empathetic responses.
Download
Paper Nr: 190
Title:

Drone Surveillance in Extreme Low Visibility Conditions
Authors:

Prachi Agrawal, Anant Verma and Pratik Narang
Abstract: Autonomous surveillance has several applications which include surveilling calamity prone areas, search and rescue operations, military operations and traffic management in smart cities. In low visibility conditions like low-light, haze, fog, snowfall, autonomous surveillance is a challenging task and current object detection models perform poorly in these conditions. Lack of datasets that capture challenging low visibility conditions is one of the reasons that limits the performance of currently available models. We propose a synthetic dataset for Human Action Recognition for search and rescue operations consisting of aerial images with different low visibility conditions including low light, haze, snowfall and also images with combinations of these low visibility conditions. We also propose a framework called ExtremeDetector for object detection in extreme low visibility conditions consisting of a degradation predictor and enhancement pool for enhancing a low visibility image and YOLOv5 for object detection in the enhanced image.
Download
Paper Nr: 193
Title:

An Empirical Heuristic Algorithm for Solving the Student-Project Allocation Problem with Ties
Authors:

Hoang H. Bach, Nguyen T. Thuong and Son T. Cao
Abstract: In this paper, we propose a simple heuristic algorithm to deal with the Student-Project Allocation problem with lecturer preferences over Projects with Ties (SPA-PT). The aim of such a problem is to find a maximum stable matching of students and projects to meet the constraints on students’ and lecturers’ preferences over projects, and the maximum numbers of students given by lecturers for each lecturer and her/his projects. Our algorithm starts from an empty matching and iteratively constructs a maximum stable matching of students and projects. For each iteration, our algorithm chooses the most ranked project of an unassigned student to assign for her/him. If the assigned project or the lecturer who offers the assigned project is over-subscribed, our algorithm removes a worth student assigned the project, where a worth student is a person corresponding to the maximum value of the proposed heuristic function. Experimental results illustrate the outperformance of the proposed algorithm w.r.t. the execution time and solution quality for solving the problem.
Download
Paper Nr: 198
Title:

Challenges in Domain-Specific Abstractive Summarization and How to Overcome Them
Authors:

Anum Afzal, Juraj Vladika, Daniel Braun and Florian Matthes
Abstract: Large Language Models work quite well with general-purpose data and many tasks in Natural Language Processing. However, they show several limitations when used for a task such as domain-specific abstractive text summarization. This paper identifies three of those limitations as research problems in the context of abstractive text summarization: 1) Quadratic complexity of transformer-based models with respect to the input text length; 2) Model Hallucination, which is a model’s ability to generate factually incorrect text; and 3) Domain Shift, which happens when the distribution of the model’s training and test corpus is not the same. Along with a discussion of the open research questions, this paper also provides an assessment of existing state-of-the-art techniques relevant to domain-specific text summarization to address the research gaps.
Download
Paper Nr: 216
Title:

A Bi-Level Genetic Algorithm to Solve the Dynamic Flexible Job Shop Scheduling Problem
Authors:

Mohamed Dhia Eddine Saouabi, Houssem Eddine Nouri and Olfa Belkahla Driss
Abstract: The dynamic flexible job shop scheduling problem (DyFJSP) is an extension of the flexible job scheduling problem (FJSP) as the production environment is characterized by a set of disturbances that require a method capable of reacting in real time in order to generate an efficient schedule in case of production failure. In this paper, we propose a bi-level genetic algorithm (BLGA) to solve the DyFJSP in order to minimize the maximum completion time (Makespan). The dynamic scenario taken into account in this work is job insertion. To evaluate the performance of our approach, we carry out experiments on Brandimarte benchmark instances. The results of the experiments show that the BLGA is characterized by its efficiency and performance in comparison with other methods published in the literature.
Download
Paper Nr: 218
Title:

Japanese Word Reordering Based on Topological Sort
Authors:

Peng Sun, Tomohiro Ohno and Shigeki Matsubara
Abstract: In Japanese, some sentences are grammatically well-formed, but not easy to read. This paper proposes a method for Japanese word reordering, which first adapts a Japanese BERT model to predict suitable pairwise orderings between two words in a sentence, and then converts the predicted results into a graph. The vertices in the graph represent the words in the sentence. The edges represent the pairwise orderings between two words. Finally, topological sort is applied to find the correct word ordering in a sentence by visiting each graph vertex. We conducted an evaluation experiment with uneasy-to-read Japanese sentences created from newspaper article sentences.
Download
Paper Nr: 231
Title:

Compression of GPS Trajectories Using Autoencoders
Authors:

Michael Kölle, Steffen Illium, Carsten Hahn, Lorenz Schauer, Johannes Hutter and Claudia Linnhoff-Popien
Abstract: The ubiquitous availability of mobile devices capable of location tracking led to a significant rise in the collection of GPS data. Several compression methods have been developed in order to reduce the amount of storage needed while keeping the important information. In this paper, we present an lstm-autoencoder based approach in order to compress and reconstruct GPS trajectories, which is evaluated on both a gaming and real-world dataset. We consider various compression ratios and trajectory lengths. The performance is compared to other trajectory compression algorithms, i.e., Douglas-Peucker. Overall, the results indicate that our approach outperforms Douglas-Peucker significantly in terms of the discrete Fréchet distance and dynamic time warping. Furthermore, by reconstructing every point lossy, the proposed methodology offers multiple advantages over traditional methods.
Download
Paper Nr: 239
Title:

Data-Driven Weather Forecast Using Deep Convolution Neural Network
Authors:

Priya Sharma, Ashish K. Patel, Pratik Shah and Soma Senroy
Abstract: Weather forecasting is an important task for the meteorological department as it has a direct impact on the day-to-day lives of people and the economy of a country. India is a diverse country in terms of geographical conditions like rivers, terrains, forests, and deserts. For the weather forecasting problem, we have taken the state of Madhya Pradesh as a case study. The current state of the art for weather forecasting is numerical weather prediction (NWP), which takes a long time and a lot of computing power to make predictions. In this paper, we have introduced a data-driven model based on a deep convolutional neural network, i.e., U-Net. The model takes weather features as input and nowcasts those features. The climate parameters considered for weather forecasting are 2m-Temperature, mean sea level pressure, surface pressure, wind velocity, model terrain height, intensity of solar radiation, and relative humidity. The model can predict weather parameters for the next 6 hours. The results are encouraging and satisfactory, given the acceptable tolerances in prediction.
Download
Paper Nr: 241
Title:

A Novel Sampling Technique for Detecting Cyber Denial of Service Attacks on the Internet of Things
Authors:

Bassam Kasasbeh and Hadeel Ahmad
Abstract: Internet of Things (IoT) devices are vulnerable to a wide range of unique security risks during the data collection and transmission processes. Due to a lack of resources, these devices increased the attack surface and made it easier for an attacker to find a target. The Denial of Service (DoS) attack is one of the most common attacks that can target all layers of the IoT protocol. Therefore, Intrusion Detection Systems (IDS) based on machine learning (ML) are the best ways to confront these risks. However, an imbalanced dataset for cyber attacks makes it difficult to detect them with ML models. We propose an undersampling technique that clusters the data set using Fuzzy C-means (FCM) and picks similar instances with the same features to ensure the integrity of the dataset. We used accuracy, precision, sensitivity, specificity, F-measure, AUC, and G-means to determine how good the results were. The proposed technique had 97.6% overall accuracy. Furthermore, it got 96.94%, 96.39%, 99.59%, 98.08%, and 97.16% True Positive Rates (TPR) in the Blackhole, Grayhole, Flooding, Scheduling, and Normal (no attacks) classes, respectively. The results show that the proposed method for detecting DoS attacks in the IoT has succeeded.
Download
Paper Nr: 250
Title:

How Far Can a Drone be Detected? A Drone-to-Drone Detection System Using Sensor Fusion
Authors:

Juann Kim, Youngseo Kim, Heeyeon Shin, Yaqin Wang and Eric T. Matson
Abstract: The recent successes of drone detection models show that leveraging the decision fusion of audio-based and vision-based features can achieve high accuracy, instead of only using unitary features. In this paper, we propose to estimate how far can a drone be detected in different distances. Drone-to-drone dataset for were collected separately using a camera and a microphone. The data are evaluated using deep learning and machine learning techniques to show how far can a drone be detected. Two different types of sensors were used for collecting acoustic-based features and vision-based features. Convolutional Neural Network (CNN) and Support Vector Machine (SVM) are utilized with audio-based features, which are Mel-Frequency Cepstral Coefficients (MFCC) and Log-Mel Spectrogram. YOLOV5 is adopted for visual feature extraction and drone detection. Ultimately, by using the sensor fusion of both domains of audio and computer vision, our proposed model achieves high performances in different distances.
Download
Paper Nr: 260
Title:

Optimising Evolution of SA-UNet for Iris Segmentation
Authors:

Mahsa Mahdinejad, Aidan Murphy, Patrick Healy and Conor Ryan
Abstract: Neuroevolution is the process of building or enhancing neural networks through the use of an evolutionary algorithm. An improved model can be defined as improving a model’s accuracy or finding a smaller model with faster training time with acceptable performance. Neural network hyper-parameter tuning is costly and time-consuming and often expert knowledge is required. In this study we investigate various methods to increase the performance of evolution, namely, epoch early stopping, using both improvement and threshold validation accuracy to stop training bad models, and removing duplicate models during the evolutionary process. Our results demonstrated the creation of a smaller model, 7:3M, with higher accuracy, 0:969, in comparison to previously published methods. We also benefit from an average time saving of 59% because of epoch optimisation and 51% from the removal of duplicated individuals, compared to our prior work.
Download
Paper Nr: 291
Title:

Normalization and Denormalization for a Document-Oriented Data
Authors:

Shady Hamouda, Mohammed Anbar and Omar Elejla
Abstract: Recently, the challenge of the increasing volume of data has led to the presentation of the “not only structured query language (NoSQL) database”. One of the most powerful types of NoSQL databases is the document-oriented database that supports a flexible schema. Normalization of the data model is one of the important research issues and there are no standard principles of normalization in the document-oriented database. Handling relationships based on normalization and denormalization has not been considered in document-oriented databases despite its importance probably because it is not recommended in creating a collection for each entity or using a reference document for all because of the need to execute a complex joint operation. Recently, many researchers have migrated from relational databases to document-oriented databases. However, their migration methods are facing issues; first is no method to normalize or de-normalize data to implement the embedded and reference document. Second, migration from a relational database to a document-oriented database does not consider how to handle various types of relationships based on normalization and de-normalization. This study proposed a way to deal with migration problems by enhancing transformation rules to map entity relational schema to document-based data schema based on normalization and denormalization data. The results of this study show that the dataset size determines whether reference or embedded documents should be used for migration.
Download
Paper Nr: 299
Title:

An Efficient Real Time Intrusion Detection System for Big Data Environment
Authors:

Faten Louati, Farah B. Ktata and Ikram Amous
Abstract: Nowadays, Security is among the most difficult issues in networks over the world. The problem becomes more challenging with the emergence of big data. Intrusion detection systems (IDSs) are among the most efficient solutions. However, traditional IDSs could not deal with big data challenges and are not able to detect attacks in real time. In this paper, a real time data preprocessing and attack detection are performed. Experiments on the well-known benchmark NSL KDD dataset show good results either in terms of accuracy rate or time of both training and testing and prove that our model outperforms other state-of-the-art solutions.
Download
Paper Nr: 304
Title:

Anomaly Detection for Multivariate Industrial Sensor Data via Decoupled Generative Adversarial Network
Authors:

Chien Wei-Chin and Wang Sheng-De
Abstract: Industrial control systems often contain sensor and actuator devices, which provide monitoring data in the form of time series, such as bridge vibrations, water distribution systems, and human physiological data. This paper proposes an anomaly detection model based on autoencoders that can consider time-series relations of the data. Moreover, the quality of the decoder output is further improved by adding a residual produced by an extra generator and discriminator. The proposed autoencoder-GAN model and detection algorithm not only improved the performance but also made the training process of GAN easier. The proposed deep learning model with the anomaly detection algorithm has been shown to achieve better results on the SWaT, BATADAL, and Rare Event Classification datasets over existing methods.
Download
Paper Nr: 306
Title:

Scenario Generation With Transitive Rules for Counterfactual Event Analysis
Authors:

Aigerim Mussina, Paulo Trigo and Sanzhar Aubakirov
Abstract: Event detection on online social networks is one of the comprehensive approaches for analyzing people’s discussions. However, it is not enough to detect an event as people often look for ways to influence the course of an event. Often, in the course of a discussion, the introduction of a new topic can shift the focus to another subject and thus move from one event to another. The causal relationship between topics and events can be explored by extracting association rules among the topics covered in each event. The scenario generation based on those causal relationships can support what-if (counterfactual) analysis and explain transitions between events. In this paper our goal is to generate what-if scenarios among topics of detected events. The association rule approach was chosen as a method for its human-readable output that can be transposed into a counterfactual scenario. We propose methods for time-window constrained topic-based what-if scenario generation founded on market-basket analysis.
Download

Area 2 - Agents

Full Papers
Paper Nr: 48
Title:

Dynamic Task Graphs for Teams in Collaborative Assembly Processes
Authors:

Ana Macedo, Liliana Antão, João Reis and Gil Gonçalves
Abstract: Collaborative robots are increasingly used in assembly processes, particularly in teams (Human-Robot or Robot-Robot), expanding the complexity and possible alternative sequences of operation and ways of team allocation to complete the assembling of a product. With this complexity, representing the possible sequences of actions needed to complete the task and the necessary constraints in a graph would improve the flexibility provided by team collaboration. However, the best sequence must be selected to increase productivity and profit, which is still challenging for a robot. This work proposes a modular system composed of three different components that, in a closed-loop interaction, allows a robotic agent to correctly plan a task given a set of operations, and optimize the task sequence allocation and scheduling plan. The effectiveness of the system is evaluated within an assembly process of different types of furniture for task sequence and allocation. The agent was able to converge successfully in three assembly scenarios: a table with 1 leg, a table with 2 legs and a table with 4 legs. Moreover, in the task allocation tests, the robotic agent was able to select actions according to the human operator expertise and its impact in the task completion time.
Download
Paper Nr: 61
Title:

Teleo-Reactive Agents in a Simulation Platform
Authors:

Vasileios Apostolidis-Afentoulis and Ilias Sakellariou
Abstract: Agent based modeling and simulation (ABMS) has been applied to a number of research areas including economics, social sciences, urban planning, epidemiology etc. with significant success. Agent simulation platforms have long been the principal tool, contributing to the wide adoption of ABMS, offering rapid model development for researchers in various fields. However, in most of the cases, agent behaviour in simulations is encoded employing standard or domain specific programming languages, with limited support for agent programming at a higher level. The present work contributes towards this direction, presenting an implementation of the Teleo-Reactive approach proposed by Nilsson for robotic control, to a well known ABMS platform, NetLogo. The aim is to allow modelers to encode complex agent models easily and thus to enhance the impact of ABMS to the respective fields.
Download
Paper Nr: 69
Title:

Autonomous Energy-Saving Behaviors with Fulfilling Requirements for Multi-Agent Cooperative Patrolling Problem
Authors:

Kohei Matsumoto, Keisuke Yoneda and Toshiharu Sugawara
Abstract: In this study, we propose a method to autonomously reduce energy consumption in the multi-agent cooperative patrol problem (MACPP) while fulfilling quality requirements. While it is crucial for the system to perform patrolling tasks as feasibly as possible, performing tasks beyond the required quality may consume unnecessary energy. First, we propose a method to reduce energy consumption by having agents individually estimate whether a given quality requirement is met through learning and consider energy-saving behaviors when diligent patrolling behavior is determined to be unnecessary. Second, we propose a method to deactivate redundant agents based on the values of parameters learned by each agent. Comparison experiments with the existing methods show that the proposed method can effectively reduce energy consumption while fulfilling the requirements. We also demonstrate that the proposed method can deactivate some agents for further energy savings.
Download
Paper Nr: 99
Title:

SNAPE: A Sequential Non-Stationary Decision Process Model for Adaptive Explanation Generation
Authors:

Amelie S. Robrecht and Stefan Kopp
Abstract: The automatic generation of explanations is an increasingly important problem in the field of Explainable AI (XAI). However, while most work looks at how complete and correct information can be extracted or how it can be presented, the success of an explanation also depends on the person the explanation is targeted at. We present an adaptive explainer model that constructs and employs a partner model to tailor explanations during the course of the interaction. The model incorporates different linguistic levels of human-like explanations in a hierarchical, sequential decision process within a non-stationary environment. The model is based on online planning (using Monte Carlo Tree Search) to solve a continuously adapted MDP for explanation action and explanation move selection. We present the model as well as first results from explanation interactions with different kinds of simulated users.
Download
Paper Nr: 146
Title:

Bottom-Up Bio-Inspired Algorithms for Optimizing Industrial Plants
Authors:

M. Umlauft, M. Gojkovic, K. Harshina and M. Schranz
Abstract: Scheduling in a production plant with a high product diversity is an NP-hard problem. In large plants, traditional optimization methods reach their limits in terms of computational time. In this paper, we use inspiration from two bio-inspired optimization algorithms, namely, the artificial bee colony (ABC) algorithm and the bat algorithm and apply them to the job shop scheduling problem. Unlike previous work using these algorithms for global optimization, we do not apply them to solutions in the solution space, though, but rather choose a bottom-up approach and apply them as literal swarm intelligence algorithms. We use the example of a semiconductor production plant and map the bees and bats to actual entities in the plant (lots, machines) using agent-based modeling using the NetLogo simulation platform. These agents then interact with each other and the environment using local rules from which the global behavior – the optimization of the industrial plant – emerges. We measure performance in comparison to a baseline algorithm using an engineered heuristics (FIFO, fill fullest batches first). Our results show that these types of algorithms, employed in a bottom-up manner, show promise of performance improvements using only low-effort local calculations.
Download
Paper Nr: 164
Title:

Fast Heuristic for Ricochet Robots
Authors:

Jan Hůla, David Adamczyk and Mikoláš Janota
Abstract: In this contribution, we describe a fast heuristic for a logical game called Ricochet Robots in which multiple robots cooperate in order to reach a goal. The heuristic recursively explores a restricted search space using subgoals that correspond to interactions of two robots. Subgoals are expanded according to an estimated length of a complete solution, which makes the algorithm reminiscent of the A* algorithm. The estimated length is a lower bound of the length of the real solution, and this allows us to prune subgoals using the best solution found thus far. After eliminating all remaining subgoals, we are guaranteed that the best solution found is the shortest solution from the restricted search space. Moreover, we show that the restricted search space contains a large portion of optimal solutions of the empirical distribution of 1 million random problems. We believe that the presented ideas should generalize to other search problems in which multiple independent agents could block or help each other.
Download
Paper Nr: 194
Title:

Studying the Impact of Transportation During Lockdown on the Spread of COVID-19 Using Agent-Based Modeling
Authors:

Shikha Bhat, Ruturaj Godse, Shruti Mestry and Vinayak Naik
Abstract: The COVID-19 pandemic has posed challenges for governments concerning lockdown policies and transportation plans. The exponential rise in infections has highlighted the importance of managing restrictions on travel. Previous research around this topic has not been able to scale and address this issue for India, given its diversity in transportation networks and population across different states. In this study, we analyze the patterns of the spread of infection, recovery, and death specifically for the state of Goa, India, for twenty-eight days. Using agent-based simulations, we explore how individuals interact and spread the disease when traveling by trains, flights, and buses in two significant settings - unrestricted and restricted local movements. Our findings indicate that trains cause the highest spread of infection within the state, followed by flights and then buses. Contrary to what may be assumed, we find that the effect of combinations of all modes of transport is not additive. With multiple modes of transport activities, the cases rise exponentially faster. We present equivalence points for the number of vehicles running per day in unrestricted and restricted movement settings, e.g., one train a day in unrestricted movement spreads the disease as eight trains a day in restricted movement.
Download
Paper Nr: 209
Title:

Swapping Physical Resources at Runtime in Embedded MultiAgent Systems
Authors:

Nilson M. Lazarin, Carlos E. Pantoja and José Viterbo
Abstract: An Embedded MultiAgent System (MAS) is a cognitive system embedded into a physical device responsible for controlling the existing resources and communicability with other devices. An Embedded MAS provides autonomy and proactivity to physical devices using the BDI model. Designing a device implies choosing sensors and actuators as resources and programming firmware and reasoning at design time. However, at runtime, resources could sometimes be damaged, presenting malfunctioning, or need to be changed. Then, performing predictive, preventive, or corrective maintenance at runtime is impossible since the designer must stop the Embedded MAS to swap resources and reprogram the system. This paper presents a novel ability for swapping resources at runtime in Embedded MAS using an extended version of Argo agents and the Jason framework. A case study analyses the new swap ability in different situations: removing and changing existing resources, adding new known and unknown resources, and causing a failure in a resource. The study case shows how the new swap ability can make devices with Embedded MAS adaptable and fault-tolerant.
Download
Paper Nr: 225
Title:

How to Find Good Coalitions to Achieve Strategic Objectives
Authors:

Angelo Ferrando and Vadim Malvone
Abstract: Alternating-time Temporal Logic (ATL) is an extension of the temporal logic CTL in which we can quantify over coalition of agents. In the model checking process, the coalitions in a given formula are fixed, so it is assumed that the user knows the specific coalitions to be checked. Unfortunately, this is not true in general. In this paper, we present an extension of MCMAS, a well-known tool that handles ATL model checking, in which we give the ability to a user to characterise the coalition quantifiers with respect to two main features: the number of agents involved in the coalitions and how to group such agents. Moreover, we give details of such extensions and provide experimental results.
Download
Paper Nr: 227
Title:

Reciprocal Adaptation Measures for Human-Agent Interaction Evaluation
Authors:

Jieyeon Woo, Catherine Pelachaud and Catherine Achard
Abstract: Recent works focus on creating socially interactive agents (SIAs) that are social, engaging, and human-like. SIA development is mainly on endowing the agent with human capacities such as communication and behavior adaptation skills. Nevertheless, the task of evaluating the agent’s quality remains as a challenge. Especially, the way of objectively evaluating human-agent interactions is not evident. To address this problem, we propose new measures to evaluate the agent’s interaction quality. This paper focuses on interlocutors’ continuous, dynamic, and reciprocal behavior adaptation during an interaction, which we refer to as reciprocal adaptation. Our reciprocal adaptation measures capture this adaptation by measuring the synchrony of behaviors including their absence of response and by assessing the behavior entrainment loop. We investigate the nonverbal adaptation, notably for smile, in dyads. Statistical analyses are conducted to improve the understanding of the adaptation phenomenon. We also studied how the presence of reciprocal adaptation may be related to different aspects of the interaction dynamics and conversational engagement. We investigate how the influence of the social dimensions of warmth and competence along with the engagement is related to reciprocal adaptation.
Download
Paper Nr: 274
Title:

A Holonic Multi-Agent Architecture For Smart Grids
Authors:

Ihab Taleb, Guillaume Guerard, Frédéric Fauberteau and Nga Nguyen
Abstract: The global warming and the increase of fossil fuel prices make the minimization of energy generation an important objective. Thus, smart grids are becoming more and more relevant in a context where we want to regulate the demand according to the available energy. This regulation can be operated thanks to Demand Side Management (DSM) tools. While different models and architectures have been developed for smart grids, only few papers used holonic architectures. For this, we propose in this paper a holonic architecture for smart grids. This type of architectures is relevant to smart grids as it allows the various actors in the grids to work even in the cases of technical problems. Holons in the proposed model are composed of five interconnecting agents that ensure flexibility on the various aspects. This model has been tested and has proven to work on 3 different scenarios. The first scenario simulates a grid in its healthy state. The second one simulates a grid where a region can be disconnected from a blackout for example. The third one simulates a grid with production mismanagement. Results show how the grid distributes the available energy depending on the available production, priorities (if any) and the assurance of the distribution across the various requesting holons.
Download
Paper Nr: 297
Title:

Differential Weight and Population Size of PRDE Traders: An Analysis of Their Impact on Market Dynamics
Authors:

George Herbert
Abstract: This paper reports results from market experiments containing Parameterised-Response Zero-Intelligence with Differential Evolution (PRDE) trader-agents. Each PRDE trader-agent in a market simultaneously uses differential evolution (DE) to adapt their trading strategy to maximise profitability. Two parameters govern the DE algorithm within each PRDE trader: the differential weight coefficient F and the number in population NP. Markets containing a homogeneous population of PRDE traders exhibit significantly different dynamics depending on the values of F and NP. The first part of this paper explores the effect that F and NP have on the profitability of markets populated by PRDE traders. The latter part of this paper proposes a new algorithm based on PRDE to maximise profitability: the Parameterised-Response Zero-Intelligence with JADE (PRJADE) trader-agent.
Download
Short Papers
Paper Nr: 2
Title:

Implicit Cooperative Learning on Distribution of Received Reward in Multi-Agent System
Authors:

Fumito Uwano
Abstract: Multi-agent reinforcement learning (MARL) makes agents cooperate with each other by reinforcement learning to achieve collective action. Generally, MARL enables agents to predict the unknown factor of other agents in reward function to achieve obtaining maximize reward cooperatively, then it is important to diminish the complexity of communication or observation between agents to achieve the cooperation, which enable it to real-world problems. By contrast, this paper proposes an implicit cooperative learning (ICL) that have an agent separate three factors of self-agent can increase, another agent can increase, and interactions influence in a reward function approximately, and estimate a reward function for self from only acquired rewards to learn cooperative policy without any communication and observation. The experiments investigate the performance of ICL and the results show that ICL outperforms the state-of-the-art method in two agents cooperation problem.
Download
Paper Nr: 25
Title:

Cross-Paradigm Interoperability Between Jadescript and Java
Authors:

Giuseppe Petrosino, Stefania Monica and Federico Bergenti
Abstract: Jadescript is a recent language for practical agent-oriented programming that aims at easing the development of multi-agent systems for real-world applications. Normally, these applications require extensive and structured reuse of existing software modules, which are commonly developed using object-oriented or legacy technologies. Jadescript has been originally designed to ease the translation to Java and, as such, it natively eases the interoperability with Java, and therefore, with mainstream and legacy technologies. This paper overviews the features that have been recently added to Jadescript to support effective two-way interoperability with Java. Moreover, this paper thoroughly discusses the main ideas behind such features by framing them in a comparison with related work, and by outlining possible directions for further developments.
Download
Paper Nr: 41
Title:

Multi-Agent Pathfinding on Large Maps Using Graph Pruning: This Way or That Way?
Authors:

Jiří Švancara, Philipp Obermeier, Matej Husár, Roman Barták and Torsten Schaub
Abstract: Multi-agent pathfinding is the task of navigating a set of agents in a shared environment from their start locations to their desired goal locations without collisions. Solving this problem optimally is a hard task and various algorithms have been devised. The algorithms can generally be split into two categories, search- and reduction-based ones. It is known that reduction-based algorithms struggle with large instances in terms of the size of the environment. A recent study tried to mitigate this drawback by pruning some vertices of the environment map. The pruning is done based on the vicinity to a shortest path of an agent. In this paper, we study the effect of choosing such shortest paths. We provide several approaches to choosing the paths and we perform an experimental study to see the effect on the runtime.
Download
Paper Nr: 51
Title:

Logic of Awareness in Agent’s Reasoning
Authors:

Yudai Kubono, Teeradaj Racharak and Satoshi Tojo
Abstract: The aim of this study is to formally express awareness for modeling practical agent communication. The notion of awareness has been proposed as a set of propositions for each agent, to which he/she pays attention, and has contributed to avoiding logical omniscience. However, when an agent guesses another agent’s knowledge states, what matters are not propositions but are accessible possible worlds. Therefore, we introduce a partition of possible worlds connected to awareness, that is an equivalence relation, to denote indistinguishable worlds. Our logic is called Awareness Logic with Partition (ALP). In this paper, we first show a running example to illustrate a practical social game. Thereafter, we introduce syntax and Kripke semantics of the logic and prove its completeness. Finally, we outline an idea to incorporate some epistemic actions with dynamic operators that change the state of awareness.
Download
Paper Nr: 63
Title:

Wall Climbing Emergent Behavior in a Swarm of Real-World Miniature Autonomous Blimps
Authors:

Tristan K. Schuler, Cameron Kabacinski, Daniel M. Lofaro, Dhawal Bhanderi, Jennifer Nguyen and Donald Sofge
Abstract: Emergent behaviors in swarms can arise when agents with simple behavioral rules produce complex group dynamics. In this work we develop both a physics-based simulation environment as well as a real-world testbed for indoor miniature autonomous blimps to analyze sensor-driven emergent behavior. During the flight experiments, the blimps had global localization from an indoor motion capture system to follow waypoints but used downward facing ultrasonic ping sensors for local-frame altitude control. After introducing a wall to the environment and commanding the swarm to the other side of the unknown obstacle, a wall climbing emergent behavior arose where the swarm agents climbed over each other as well as the wall to reach the goal. We demonstrate how modifying the sensor characteristics between trials and changing the swarm size affects this emergent behavior in both simulation and with real-world blimps, and compare the results.
Download
Paper Nr: 101
Title:

A Novel Group-Based Firefly Algorithm with Adaptive Intensity Behaviour
Authors:

Adam Robson, Kamlesh Mistry and Wai L. Woo
Abstract: This paper presents novel modifications to the Firefly Algorithm (FA) that manipulate the functionality of the intensity and attractiveness of fireflies through the incorporation of grouping behaviours into the movement of the fireflies. FA is one of the most well-known and actively researched swarm-based algorithms, gaining notoriety for the powerful search capability offered and overall computational simplicity. While the FA is an effective optimisation algorithm, it is unfortunately susceptible to the issue of premature convergence and oscillations within the swarm, which can lead to suboptimal performance. In the original FA formulation, at each iteration fireflies will instinctively move towards the most intensely bright firefly which is in closest proximity to them. The algorithm proposed in this paper manipulates the movement of the fireflies through modification of this intensity and attraction relationship, allowing the swarm to move in different ways, ultimately increasing the search diversity within the swarm. While group-based FAs have been proposed previously, the group-based FAs presented in this paper utilise a different approach to creating groups, implementing groupings based upon firefly performance at each iteration, resulting in continually varying groupings of fireflies, to further increase search diversity and maintain computational simplicity.
Download
Paper Nr: 112
Title:

Measuring Emotion Velocity for Resemblance in Neural Network Facial Animation Controllers and Their Emotion Corpora
Authors:

Sheldon Schiffer
Abstract: Single-actor facial emotion video corpora for training NN animation controllers allows for a workflow where game designers and actors can use their character performance training to significantly contribute to the authorial process of animated character behaviour. These efforts result in the creation of scripted and structured video samples for a corpus. But what are the measurable techniques to determine if a corpus sample collection or NN design adequately simulates an actor’s character for creating autonomous emotion-derived animation? This study focuses on the expression velocity of the predictive data generated by a NN animation controller and compares it to the expression velocity recorded in the ground truth performance of the eliciting actor’s test data. We analyse four targeted emotion labels to determine their statistical resemblance based on our proposed workflow and NN design. Our results show that statistical resemblance can be used to evaluate the accuracy of corpora and NN designs.
Download
Paper Nr: 116
Title:

Conflicting Moral Codes for Self-Driving Cars: The Single Car Case
Authors:

Ahmed Ibrahim, Yasmin Mansy, Nourhan Ehab and Amr Elmougy
Abstract: There has been incremental advances in the development of self-driving cars. However, there are still several gaps that must be filled before a fully autonomous vehicle can be achieved. The ability to resolve conflicts in the event of an unavoidable accident is one of the most prominent and crucial aspects of a self-driving car that is currently absent. To address this gap, this paper aims to resolve moral conflicts in self-driving cars in the case of an unavoidable accidents. Assuming we have a predefined rule set that specifies how a car should morally react, any clash between these rules could result in a critical conflict. In this paper, we propose a novel procedure to resolve such conflicts by combining the Thomas Kilmann conflict resolution model together with decision trees. Evaluation results showcase that our proposed procedure excels in distinct ways, enabling the self-driving car to make a decision that will yield the best moral outcome in conflicting scenarios.
Download
Paper Nr: 118
Title:

Machine Learning for Cognitive BDI Agents: A Compact Survey
Authors:

Ömer I. Erduran
Abstract: The concept of Cognitive Agents has its roots in the early stages of Multi-Agent Systems research. At that time, the understanding of the term Agent was referring to Software Agents with basic capabilities of perception and action in a proper environment adding potential cognitive capabilities inside the agent architecture. A fundamental drawback of the concept is the barrier of learning new capabilities since the full properties of the agent are hard coded. Over the years, research in Agent-Oriented Programming has provided interesting approaches with promising results in the interplay between Machine Learning methods and Cognitive Agents. Such a combination is realized by an integration process of Machine Learning algorithms into the agent cycle in the specific architecture. This survey is a review of combining both, Machine Learning and BDI Agents as a selected form of Software Agent, including the applied concepts and architectures for different scenarios. A categorization scheme named ML-COG is introduced to illustrate the integration perspectives for both paradigms. The reviewed literature is then assigned to this scheme. Finally, a selection of relevant research questions and research gaps is presented as worthwhile to be investigated.
Download
Paper Nr: 130
Title:

A Study Toward Multi-Objective Multiagent Reinforcement Learning Considering Worst Case and Fairness Among Agents
Authors:

Toshihiro Matsui
Abstract: Multiagent reinforcement learning has been studied as a fundamental approach to empirically optimize the policies of cooperative/competitive agents. A previous study proposed an extended class of multi-objective reinforcement learning whose objectives correspond to individual agents, and the worst case and fairness among the objectives was considered. However, that work concentrated on the case of joint-state-action space that is handled by a centralized learner performing an offline learning. Toward decentralized solution methods, we investigate the situations including on-line learning where agents individually own their learning tables and selects optimum joint actions by cooperatively combining the decomposed tables with other agents. We experimentally investigate the possibility and influence of the decomposed approach.
Download
Paper Nr: 157
Title:

Hierarchical Constraint Logic Programming for Multi-Agent Systems
Authors:

Hiroshi Hosobe and Ken Satoh
Abstract: Logic programming is a powerful tool for modeling and processing multi-agent systems (MASs), where problems are collaboratively solved by multiple agents that exchange messages. Especially, (constraint) logic programming with default rules is useful for speculative computation for MASs, where tentative solutions are computed before answers arrive from other agents. However, the previous frameworks became complicated to enable speculative computation for MASs. In this paper, we propose a framework using hierarchical constraint logic programming (HCLP) for speculative computation for MASs. HCLP is an extension of constraint logic programming that allows soft constraints with hierarchical preferences. We simplify our MAS framework by utilizing HCLP and soft constraints to handle MASs with default rules. We show a prototype implementation of our framework and a case study on its execution.
Download
Paper Nr: 180
Title:

Parallel and Distributed Epirust: Towards Billion-Scale Agent-Based Epidemic Simulations
Authors:

Sapana Kale, Shabbir Bawaji, Akshay Dewan, Meenakshi Dhanani, Kritika Gupta, Harshal Hayatnagarkar, Swapnil Khandekar, Jayanta Kshirsagar, Gautam Menon and Saurabh Mookherjee
Abstract: EpiRust is an open source, large-scale agent-based epidemiological simulation framework developed using Rust language. It has been developed with three key factors in mind namely 1. Robustness, 2. Flexibility, and 3. Performance. We could demonstrate EpiRust scaling up to a few millions of agents, for example a COVID19 infection spreading through Pune city with its 3.2 million population. Our goal is to simulate larger cities like Mumbai (with 12 million population) first, and then entire India with its 1+ billion population. However, the current implementation is not scalable for this purpose, since it has a well-tuned serial implementation at its core. In this paper, we share our ongoing journey of developing it as a highly scalable cloud ready parallel and distributed implementation to simulate up to 100 million agents. We demonstrate performance improvement for Pune and Mumbai cities with 3.2 and 12 million populations respectively. In addition, we discuss challenges in simulating 100 million agents.
Download
Paper Nr: 204
Title:

Modelling Agents in Industry 4.0 Applications Using Asset Administration Shell
Authors:

Nikoletta Nikolova and Sjoerd Rongen
Abstract: Within Industry 4.0, more applications include multi-agent systems and integrated software agents in the newly developed solutions, as they can provide a valuable contribution to the manufacturing processes. This makes it important to create a digital representation of those virtual assets, similar to how this is done for physical ones. The Asset Administration Shell (AAS) has been designed for just this purpose - to create models of assets, containing all the relevant information. Currently, this is standard for physical assets, however, it could be of value to extend it beyond. We propose the usage of the AAS for creating information models of software agents and suggest a generic approach to apply the AAS meta-model to ensure semantic interoperability between them. For this purpose, we outline a structure and a set of specific submodels to group agent attributes, which can provide a description of all relevant information for a given task. We provide two examples of concrete agents and outline how this approach will be further validated within future use cases.
Download
Paper Nr: 211
Title:

Study on Decentralized Anytime Evolutionary Algorithm for DCOPs Containing Adversarial Agents
Authors:

Toshihiro Matsui
Abstract: The Distributed Constraint Optimization Problem (DCOP) is a fundamental optimization problem that represents the cooperation of multiple agents. An extended class of DCOPs contains potentially adversarial agents that can select arbitrary decisions or the worst one, and the goal is to find a safe solution under the worst case by emulating adversarial agents. Such problems are important for addressing risky situations in real world applications. Although several exact solution methods based on distributed asynchronous game-tree search for the case have been studied, their scalability is limited by the tree-width of constraint graphs that represent the DCOPs. We study the application of decentralized optimization methods based on an anytime evolutionary algorithm for DCOPs to the cases containing adversarial agents. We employ solution methods to minimize upper bound cost values, investigate several heuristic unbounded methods, and experimentally evaluate our proposed approach.
Download
Paper Nr: 222
Title:

A Game Theoretic Approach to Attack Graphs
Authors:

Davide Catta, Antonio Di Stasio, Jean Leneutre, Vadim Malvone and Aniello Murano
Abstract: An attack graph is a succinct representation of all the paths in an open system that allow an attacker to enter a forbidden state (e.g., a resource), besides any attempt of the system to prevent it. Checking system vulnerability amounts to verifying whether such paths exist. In this paper we reason about attack graphs by means of a game-theoretic approach. Precisely, we introduce a suitable game model to represent the interaction between the system and the attacker and an automata-based solution to show the absence of vulnerability.
Download
Paper Nr: 237
Title:

Automated Agent Migration over Distributed Data Structures
Authors:

Vishnu Mohan, Anirudh Potturi and Munehiro Fukuda
Abstract: In contrast to conventional data streaming, we take an agent-based approach where a large number of reactive agents collaboratively analyze attributes or shapes of distributed data structures that are mapped over a cluster system. Our approach benefits distributed graph database and GIS as agents are dispatched to data of interest and navigate over nearby structured data for further exploration and exploitation. The successful key to this approach is how to code agent propagation, forking, and flocking over data structures. We automated such agent migration in our MASS (multi-agent spatial simulation) library and wrote four benchmark programs with these migration functions. The benchmarks include breadth-first search, triangle counting, range search, and closet pair of points in space. This paper demonstrates improvements of parallel performance with the automated migration and presents our programmability comparison with Repast Simphony.
Download
Paper Nr: 242
Title:

Coupled Assignment Strategy of Agents in Many Targets Environment
Authors:

Azizkhon Afzalov, Ahmad Lotfi and Jun He
Abstract: There are multi-agent algorithms that provide solutions with the shortest path without considering other pursuing agents. However, less attention has been paid to computing an assignment strategy for the pursuers that assign targets before the move action. Besides, the pathfinding problem for multiple agents becomes even more challenging if the goal destinations change over time. The path-planning problem for multiple pursuing agents requires more efficient assignment strategy algorithms. Therefore, this study considers existing and the most recent solutions and conducts experiments in a dynamic environment where multiple pursuing agents are outnumbered and required to capture the moving targets for a successful outcome. The existing cost function strategies, such as sum-of-costs and makes, are compared and analysed to the recent twin cost, cover cost and weighted cost assignment strategies. The results indicate that the recent criterion, the cover cost algorithm, shows the optimal outcomes in terms of pathfinding cost and runtime. Statistical analyses have also demonstrated the significance of the findings.
Download
Paper Nr: 259
Title:

The Missing Tip: Lack of Micro-Movements Impairs Navigation Realism in Artificial Social Agents
Authors:

Jacob Sharp and Ulysses Bernardet
Abstract: Navigation is critical to an intelligent social agent’s ability to interact with the world and any other agent, virtual or otherwise. In order to create a truly realistic artificial social agent, unconscious human micro-movements need to be simulated. We see this as an important goal for the research area. Examples of these micro-movements include orienting while walking and back-stepping, strafing with attention focused elsewhere, and micro-orientations during locomotion. We postulate that there is a gap in research around these micro-movements within the field of navigation that we hope to contribute to filling. Most research in this field is focused on the understandably important pathfinding aspect of navigation; moving between two spatial locations. There is little to no research being done on micro-movements and making a truly realistic navigation system for artificial social agents. Moreover, there exists no canonical way of describing these movements and ”micro-movements” that are so characteristic for human spatial behaviour. Here we propose a set of standardised descriptors of movement configurations, that will be able to be used as building blocks for spatial behaviour experimentation, and as the basis for behaviour generation models. We see this as an important tool in the creation of navigation systems that are able to more readily include these kinds of behaviours, with hope that the aforementioned configurations will improve development of realistic movement systems.
Download
Paper Nr: 261
Title:

REPMAS: A Requirements Engineering Process for MultiAgent Systems: An Application Example
Authors:

Iderli S. Filho, Gilleanes A. Guedes and Giovane D. Mendonça
Abstract: Requirements engineering is a crucial phase for the development process of any software, including multi-agent systems. This particular kind of software is composed of agents, autonomous and proactive entities which can collaborate among themselves to achieve a given goal. However, multi-agent systems have some particular requirements that are not normally found in other software, making the requirements engineering general processes and techniques less efficient. Taking this into account, this work presents a specific requirements engineering process for multi-agent systems with emphasis in a consolidated model in the cognitive agents development (BDI Model). This process supports the requirements engineering subareas of elicitation, analysis, specification, and validation, thus, presenting a wide coverage of the requirements engineering area. The proposed process was evaluated through its application in requirements engineering of Heraclito multi-agent system. This assessment allowed us to identify future works and improvement points in our work.
Download
Paper Nr: 264
Title:

Towards Modelling and Verification of Social Explainable AI
Authors:

Damian Kurpiewski, Wojciech Jamroga and Teofil Sidoruk
Abstract: Social Explainable AI (SAI) is a new direction in artificial intelligence that emphasises decentralisation, transparency, social context, and focus on the human users. SAI research is still at an early stage. Consequently, it concentrates on delivering the intended functionalities, but largely ignores the possibility of unwelcome behaviours due to malicious or erroneous activity. We propose that, in order to capture the breadth of relevant aspects, one can use models and logics of strategic ability, that have been developed in multi-agent systems. Using the STV model checker, we take the first step towards the formal modelling and verification of SAI environments, in particular of their resistance to various types of attacks by compromised AI modules.
Download
Paper Nr: 267
Title:

Strategy Analysis for Competitive Bilateral Multi-Issue Negotiation
Authors:

Takuma Oishi and Koji Hasebe
Abstract: In most existing negotiation models, each agent aims only to maximize its own utility, regardless of the utility of the opponent. However, in reality, there are many negotiations in which the goal is to maximize the relative difference between one’s own utility and that of the opponent, which can be regarded as a kind of zero-sum game. The objective of this study is to present a model of competitive bilateral multi-issue negotiation and to analyze strategies for negotiations of this type. The strategy we propose is that the agent makes predictions both about the opponent’s preference and how the opponent is currently predicting its own preference. Based on these predictions, the offer that the opponent is most likely to accept is proposed. To demonstrate the usefulness of this strategy, we conducted experiments in which agents with several strategies, including ours, negotiated with one another. The results demonstrated that our proposed strategy had the highest average utility and winning rate regardless of the error rate of the preference prediction.
Download
Paper Nr: 3
Title:

Online HTN Planning for Data Transfer and Utilization Considering Legal and Ethical Norms: Case Study
Authors:

Hisashi Hayashi and Ken Satoh
Abstract: Data transfer among servers is crucial for distributed data mining because many databases are distributed around the world. However, as data privacy is becoming more legally and ethically protected, it is necessary to abide by the laws and respect the ethical guidelines when transferring and utilizing data. Because information affecting legal/ethical decision-making is often distributed, the data-transfer plan must be updated online when new information is obtained while transferring data among servers. In this study, we propose a dynamic hierarchical task network (HTN) planning method that considers legal and ethical norms while planning multihop data transfers and data analyses/transformations. In our knowledge representation, we show that datatransfer tasks can be represented by the task-decomposition rules of total-order HTN planning. We also show that legal norms can be expressed as the preconditions of tasks and actions, and ethical norms can be expressed as the costs of tasks and actions where legal norms cannot be violated, but ethical norms can be violated if necessary following the ethical theory of utilitarianism. In the middle of the plan execution, the online planner dynamically updates the plan based on new information obtained in accordance with laws and ethical guidelines.
Download
Paper Nr: 29
Title:

Operational Semantic of an AgentSpeak(L) Interpreter Using Late Bindings
Authors:

Frantisek Vidensky, Frantisek Zboril, Radek Koci and Frantisek V. Zboril
Abstract: Although BDI systems have long been studied in the field of agent-based programming, there are still problems open for research. One problem is that some parts of systems are non-deterministic in the original specification. However, finding a suitable deterministic method can lead to improved rationality of an agent’s behaviour. In our previous work, we introduced late binding into the interpretation of AgentSpeak(L) language. The main benefit of this approach is that the interpreter chooses substitutions only when needed, thus avoiding unnecessary and incorrect substitution selection. In this paper, we present a formal operational semantics for an interpreter using late binding variables. A well-specified operational semantics is necessary for the implementation of such an interpreter and its further development.
Download
Paper Nr: 37
Title:

Analyzing the Effectiveness of Stereotypes
Authors:

Nanda K. Sreenivas and Robin Cohen
Abstract: In this paper, we present an agent-based model to understand the effectiveness of stereotpyes in supporting decision making of users in online settings such as e-marketplaces. We define four different agents types that combine information from stereotypes and past interactions in different ways. Through simulation, we find that agents that use their memory primarily and stereotypes as a last resort perform the best. We discuss how this work relates to that of trust modeling in multiagent systems and reflect briefly on how the emotional makeup of a user may influence what is best for decisions about stereotype use.
Download
Paper Nr: 38
Title:

Virtual Agent Behavior Modeling in Case of a Risky Situation in a Virtual Electrical Substation
Authors:

Dilyana Budakova, Velyo Vasilev, Lyudmil Dakovski and Stanimir Stefanov
Abstract: In this paper, the behavior of a realistically represented intelligent virtual agent (IVA) that accompanies students during their visit to a virtual electrical substation is modeled. The choice of technologies for modeling the agent and the task environment is considered. The properties of the task environment are discussed. The agent’s behavior when a risky situation occurs is investigated. For this purpose, an IVA behavior model, based on psychological theories of motivation, emotions, and power is proposed. A change in the IVA priorities and, as a consequence, a change in its goal is modeled. Results of a survey, studying the trust, which the IVA receives from the students, are presented. To have a more realistic IVA, the model includes knowledge of the environment, the shortest evacuation route learning, visitor training locations, priorities, emotions, social power strategy, set goals, abilities to learn, abilities to change priorities and goals when a risk occurs, and a role of a specialist – electrician.
Download
Paper Nr: 62
Title:

Ontology-Based Solution for Building an Intelligent Searching System on Traffic Law Documents
Authors:

Vuong T. Pham, Hien D. Nguyen, Thinh Le, Binh Nguyen and Hung Q. Ngo
Abstract: In this paper, an ontology-based approach is used to organize the knowledge base of legal documents in road traffic law. This knowledge model is built by the improvement of ontology Rela-model. In addition, several searching problems on traffic law are proposed and solved based on the legal knowledge base. The intelligent search system on Vietnam road traffic law is constructed by applying the method. The searching system can help users to find concepts and definitions in road traffic law. Moreover, it can also determine penalties and fines for violations in the traffic. The experiment results show that the system is efficient for users’ typical searching and is emerging for usage in the real-world.
Download
Paper Nr: 147
Title:

Towards a Fleet of Autonomous Haul-Dump Vehicles in Hybrid Mines
Authors:

Alexander Ferrein, Michael Reke, Ingrid Scholl, Benjamin Decker, Nicolas Limpert, Gjorgji Nikolovski and Stefan Schiffer
Abstract: Like many industries, the mining industry is facing major transformations towards more sustainable and decarbonised operations with smaller environmental footprints. Even though the mining industry, in general, is quite conservative, key drivers for future developments are digitalisation and automation. Another direction forward is to mine deeper and reduce the mine footprint at the surface. This leads to so-called hybrid mines, where part of the operation is open pit, and part of the mining takes place underground. In this paper, we present our approach to running a fleet of autonomous hauling vehicles suitable for hybrid mining operations. We present a ROS 2-based architecture for running the vehicles. The fleet of currently three vehicles is controlled by a SHOP3-based planner which dispatches missions to the vehicles. The basic actions of the vehicles are realised as behaviour trees in ROS 2. We used a deep learning network for detection and classification of mining objects trained with a mixing of synthetic and real world training images. In a life-long mapping approach, we define lanelets and show their integration into HD maps. We demonstrate a proof-of-concept of the vehicles in operation in simulation and in real-world experiments in a gravel pit.
Download
Paper Nr: 192
Title:

A Study of the Effectiveness of English Speaking of Teachable Agent using AI Chatbot
Authors:

Kyung A. Lee, Soon-Bum Lim and Shankara N. Nagarajan
Abstract: In an environment where English is a foreign language (English as a foreign language: EFL), English learners use AI voice chatbots for English-speaking practice activities. They enhance their speaking motivation and provide opportunities for communication practice, ultimately leading to English language learning. It can improve their speaking skills. However, if they are preschoolers or elementary school students with no experience learning English, a conversation may not be possible using the AI voice chatbot system. In this study, we propose a teachable agent using an AI voice chatbot that can be easily used even for elementary school students and can enhance the learning effect. The existing Teachable Agent is a method that makes inferences with the knowledge acquired from the learner and answers questions using a path search algorithm. However, applying the Teachable Agent system to language learning is complex, an activity based on tense, context, and memory. This paper proposed a new TA method by reflecting the learner’s English pronunciation and level to the teachable agent and generating the agent’s answer according to the learner’s error. Moreover, a teachable Agent AI chatbot prototype was implemented with an AI voice chatbot.
Download
Paper Nr: 208
Title:

Towards User-Centred Validation and Calibration of Agent-Based Models
Authors:

Etienne Tack, Gilles Énée, Thomas Gaillard, Jean-Marie Fotsing and Frédéric Flouvat
Abstract: This paper describes a path to a user-centred approach for calibration and validation of agent-based models, particularly for spatially explicit models. Including the end-user in these critic modelling steps, we hope for better models that converge more easily toward reality. Using experts’ knowledge, validation measures and feedback links to model parameters can be established. However, experts are not necessarily proficient in computer science. Tools should be created to help the transmission of their knowledge. With this paper, complying with a user-centred approach, we suggest using user-defined validation measures and a visual programming language to let the experts adjust themselves the behaviour rules of the agents.
Download
Paper Nr: 210
Title:

A Normative Multiagent Approach to Represent Data Regulation Concerns
Authors:

Paulo H. Alves, Fernando A. Correia, Isabella Z. Frajhof, Clarisse Sieckenius de Souza and Helio Lopes
Abstract: Data protection regulation is crucial to establishing the appropriate conduct in sharing and maintaining personal data. It aims to protect the Data Subjects’ data, and to define Data Controllers’ and Processors’ obligations. However, modeling systems to represent and comply with those regulations can be challenging. In this sense, Multiagent System (MAS) presents an opportunity to overcome this challenge. MAS is an artificial intelligence approach that enables the simulation of independent software agents considering environmental variables. Thus, combining data regulation directives and Normative MAS (NMAS) can allow the development of systems among distinct data regulation jurisdictions properly. This work proposes the DR-NMAS (Data Regulation by NMAS) employing Adaptative Normative Agent - Modeling Language (ANA-ML) and a Normative Agent Java Simulation (JSAN) extension to address data regulation concerns in an NMAS. As a result, we present a use case scenario in the Open Banking domain to employ the proposed extensions. Finally, this work concludes that NMAS can represent data regulation modeling and its application.
Download
Paper Nr: 228
Title:

Learning to Participate Through Trading of Reward Shares
Authors:

Michael Kölle, Tim Matheis, Philipp Altmann and Kyrill Schmid
Abstract: Enabling autonomous agents to act cooperatively is an important step to integrate artificial intelligence in our daily lives. While some methods seek to stimulate cooperation by letting agents give rewards to others, in this paper we propose a method inspired by the stock market, where agents have the opportunity to participate in other agents’ returns by acquiring reward shares. Intuitively, an agent may learn to act according to the common interest when being directly affected by the other agents’ rewards. The empirical results of the tested general-sum Markov games show that this mechanism promotes cooperative policies among independently trained agents in social dilemma situations. Moreover, as demonstrated in a temporally and spatially extended domain, participation can lead to the development of roles and the division of subtasks between the agents.
Download
Paper Nr: 269
Title:

An Agent-Based Model Study on Subsidy Fraud Risk in Technological Transition
Authors:

Hao Yang, Xifeng Wu and Yu Chen
Abstract: In modern society, government subsidy policies play a pivotal role in developing new technologies. Although subsidy policies have a long history, the resulting subsidy fraud problem consumes social resources and hinders the development of new technologies. In this paper, we attempt to derive the factors affecting the risk of performing the subsidy fraud based on a validated agent-based model for technological transition. We first review the literature on subsidies and the definition of subsidy policies. We perform a mathematical analysis of the agent-based model and calculate the critical value for subsidy rates, which may cause a dramatic change in the probability of subsidy fraud to occur. We conducted a series of numerical experiments to show the validity of the critical subsidy rates. And we also correlates and classifies three scenarios between the situation of technology diffusion and development and the risk of subsidy fraud. Finally, the causal factors of subsidy fraud are examined by analyzing the various stakeholders involved in the subsidy fraud in the actual situation.
Download