Abstracts Track 2023

Area 1 - Artificial Intelligence
Area 2 - Agents

Area 1 - Artificial Intelligence

Nr: 4
Title:

AdaPipe: A Recommender System for Adaptive Computation Pipelines in Cyber-Manufacturing Computation Services
Authors:

Ran Jin and Xiaoyu Chen
Abstract: The industrial cyber-physical systems will accelerate the transformation of offline data-driven modeling to fast computation services, such as computation pipelines for prediction, monitoring, prognosis, diagnosis, and control in factories. However, it is computationally intensive to adapt computation pipelines to heterogeneous contexts in manufacturing. In this paper, we propose to rank and select the best computation pipelines to match contexts and formulate the problem as a recommendation problem. The proposed method Adaptive computation Pipelines (AdaPipe) considers similarities of computation pipelines from word embedding, and features of contexts. Thus, without exploring all computation pipelines extensively in a trial-and-error manner, AdaPipe efficiently identifies top-ranked computation pipelines. We validated the proposed method with 60 bootstrapped data sets from three real manufacturing processes: thermal spray coating, printed electronics, and additive manufacturing. The results indicate that the proposed recommendation method outperforms traditional matrix completion, tensor regression methods, and a state-of-the-art personalized recommendation model. See Details of the Published Paper: https://doi.org/10.1109/TII.2020.3035524
Nr: 13
Title:

Korean Speech Emotion Recognition Based on Deep Learning Model with Two-stream BiLSTM and CNN Structure
Authors:

Keun-Chang Kwak and A-Hyeon Jo
Abstract: Understanding a person's emotional state is an important factor in communicating with people. Among them, speech is a means to express emotions easily and naturally. Speech emotion recognition technology is an important technology in human computer interaction (HCI) systems, and it is important to accurately identify emotions. Accordingly, in this paper, we propose a two-stream based emotion recognition model of Bidirectional Long-Short Term Memory (Bi-LSTM) and Convolutional Neural Networks (CNN) from the Korean speech emotion database. The data used in the experiment is the Korean speech emotion recognition database built directly by Chosun University. This database was built for 200 people and was composed of voice recording files for a total of 8 emotions (happy, neutral, angry, sad, chagrin, disgust, fear, and surprise). Participants attached SONY ECM-CS3 stereo pin-in microphones to their collars and recorded in a noise-free environment as quiet as possible considering the characteristics of voice data that are sensitive to noise. Acquisition of data was carried out by setting a situation suitable for each emotion and a short sentence suitable for it, and acting it with emotion. Speech files were recorded at 48,000 Hz and saved in wav format. This database includes 80 files for each participant, 10 for each emotion, and thus includes a total of 16,000 voice emotion data, 2,000 for each emotion. The deep learning model is designed by an emotion recognition model by combining Bi-LSTM and YAMNet, a CNN-based transfer learning model with 2-stream structure. To compare the recognition performance, we perform the speech feature extraction method and the deep learning model in several cases. As a result, the speech emotion recognition performance of Bi-LSTM or YAMNet single model was 90.38% and 94.91%, respectively. However, when two models are combined in 2-stream rather than a single model, the performance of the model is 96%, which proves that the minimum 1.09% to maximum 5.62% improvement over the single model.
Nr: 16
Title:

Robust Adversarial Mixup Training for Reliable Deep Neural Networks to Achieve Safe AI
Authors:

Kyungpil Gwon, Sangmoon Lee, Ju H. Park and Joonhyuk Yoo
Abstract: Detecting abnormal data whose distribution is far from the nominal distribution of training data and a generalization ability to flexibly respond to naturally or artificially occurring corruption phenomena are essential to achieving reliable results in various machine learning applications. However, Deep Neural Networks (DNNs) have recently been found to exhibit the over-confidence problem in Out-of-Distribution (OOD) data such as abnormal samples and a reliability issue misclassifying some data disturbed by noise to wrong classes. To address these problems, this paper proposes a novel unified framework for OOD detection and OOD generalization called by Adversarial Mixup (AM) training. The proposed training method makes an augmentation effect on the data to increase the generalization performance and exploits the Mahalanobis distance by estimating a statistical distance between data clusters through stochastic modeling of data distributions in the inference process of DNNs, which effectively solves the reliability problems related to the over-confidence due to the softmax function. The proposed AM method simultaneously addresses both OOD detection and OOD generalization issues by providing more efficient reliability and robustness enhancement strategy compared with conventional methods focusing only on performance improvement of DNNs. Experimental evaluation shows that the proposed technique induces low confidence scores for OOD data and increases the OOD generalization performance for corrupted data, increasing performance up to 74% compared to that achieved by previous SOTA methods.
Nr: 18
Title:

Suggestion of Model for Modifying Positioning of Football by Genetic Algorithm
Authors:

Yuya Jingushi, Yuichi Sei, Yasuyuki Tahara and Akihiko Ohsuga
Abstract: In recent years, researchers have studied soccer with the development of computers and analytical methods. Players’ positioning has become increasingly important, but it’s difficult to interpret positioning because correct positioning varies greatly depending on the situation. Pitch control is one of the most well-known research topics in positioning. It can visualize pitch domination. If the defenders adjust their positioning in a phase for decreasing the areas which are controlled by attackers, they can decrease the risk of conceding a goal. However, there are areas where the attacker has the advantage but the probability of goals or assists is low. Changing the positioning based on only Pitch Control in such areas is difficult. Therefore, in this study, we introduce a method for quantifying the risk of conceding a goal and propose a model to output the modification of positioning in a phase based on the method. For quantifying the positioning, we consider the ease of making goals and assists in addition to pitch domination. Pitch domination is expressed using Pitch Control. The ease of conceding a goal is shown by xG, which represents the expected value of goals, and the newly proposed SxA, which represents the expected value of assists. For each grid of a 1m x 1m segment of a 106m x 68m pitch, Pitch Control, xG and SxA are calculated. To quantify the positioning, we propose the formula “Pitch Control × ( xG ＋ SxA × α )”. For making a model modifying the positioning, we use Genetic Algorithm. The movement of a player is represented by two parameters: distance and direction. The set of 22 parameters summarized for the movement of the 11 players in the defending team, was defined as the chromosomes of the individuals. The distance was set to be between 0 m and 2 m, as a significant movement of the players was not feasible. In addition, the method of quantifying the positioning mentioned above was adopted for the evaluation function. For the quantification of positioning, we compared the outputs with a summary of areas that 7 people who had learned to play soccer answered. The result showed that the areas of the output and the areas answered are consistent in some areas. It is in the area where the attacking player is running in front of the goal and it is likely to lead to an assist besides the goal. On the other hand, areas in the opposite direction of the players' directions and movements were not considered highly dangerous in the summary. For the positioning modification model, we compared the outputs with the results of random generation at the same number of individuals as generated by the genetic algorithm. As a result, the evaluation value improved by about 4%. In addition, an improvement linked to the players was observed by having 11 players moving at the same time. Continued improvement of the model would help many people interpret positioning.
Nr: 19
Title:

Predicting Weighing Deviations in the Dispatch Workflow Process: A Case Study in a Cement Industry
Authors:

Julio L. de Barros, Francisco Cunha, Cândido Martins, Paulo Pedrosa and Paulo Cortez
Abstract: The supply chain (SC) consists of a complex and unique network, which includes several entities, processes and resources. Logistics constitutes one of the crucial factors in the success of the SC, which consists of planning and coordinating the movement of products in a timely, safe and effective way. Logistics management activities comprise inbound and outbound transportation management, fleet management, warehousing, materials handling, order fulfilment, logistics network design, and inventory management, among others. The order fulfilment process (OFP) consists of one of the key business processes for the organization SC and represents a core process for the operational logistics flow. Indeed, the problems associated with this process can result in high losses, such as losses of customer’s confidence, reputation, and revenues. The assessment of the OFP problems and the creation of digital solutions, according to the Industry 4.0 concept, to support this process allow the improvement of the efficiency of the supply chain and consequently improves the organizational performance and achieve a competitive advantage. The dispatch workflow process is an integral part of the OFP and is also a crucial process in the SC of cement industry organizations. In this work, we focus on enhancing the order fulfilment process by improving the dispatch workflow process, specifically with respect to the cement loading process. Thus, we proposed a machine learning (ML) approach to predict weighing deviations in the cement loading process motivated mainly by (i) the nonexistence of scientific works that focus on improving OFP, more specifically regarding the problem of deviation in weighings during the dispatch workflow process, (ii) the lack of studies regarding the cement industry supply chain, although the SCs problems are an attractive topic and, (iii) the scarce consideration of ML techniques in SC management scientific studies. Hence, the occurrence of weight deviations, which represents an anomaly in the loading of cement bags, poses a complex problem that directly impacts the OFP and consequently the SC performance, resulting in several losses, including monetary and service level losses. Indeed, in this work, we adopted a realistic and robust rolling window (RW) scheme to evaluate six classification models, namely decision tree (DT), random forest (RF), support vector machines (SVM), gradient-boosted tree (GBT), extreme gradient-boosting tree (XGBT), and multilayer perceptron (MLP) in a real-world case study in the company Cachapuz - Weighing & Logistics Systems, Lda, Portugal, from which the random forest (RF) model provides the best predictive performance with a median AUC of 0.937 over the twenty iterations of the RW, followed by the GBT, XGBT, SVM, MLP, and DT models. We also extracted explainable knowledge from the RF classifier by using the Shapley additive explanations (SHAP) method, demonstrating the influence of each input data attribute used in the prediction process. Hence, we provide the top ranking of features and demonstrate that eight of the eleven features considered were the result of feature engineering. The selected model (RF model) is deployed on a technological architecture that we propose, which enables gathering data from the existing Cachapuz database, training the model and making predictions from that data, and presenting the model outcomes on a PowerBI dashboard, together with useful explanations of the predictions.

Area 2 - Agents

Nr: 17
Title:

Towards an Hierarchical Model-Based Reinforcement Learning Approach to Dynamic Decision-Making in Uncertain Environment
Authors:

Annika Österdiekhoff, Nils Heinrich, Nele Russwinkel and Stefan Kopp
Abstract: Dynamic decision-making as the ability to choose between options in a constantly changing environment in real time, is easily applied by humans who use cognitive abilities like goal-directed planning (Eppe et al., 2022). Autonomous intelligent agents often fail to act in these complex dynamic environments. Currently, the most prominent approach to model the learning of robust action policies in artificial autonomous agents in dynamic environments is reinforcement learning (RL). However, such models do not perform decision-time planning but follow a reactive policy that has been optimized over a vast number of trials. The indeterminacies of uncertain environments might not be fully explored during training. We argue that by including goal-directed planning, i.e., simulating various decision options, autonomous agents will improve their ability to adapt to these environments. One way of implementing goal-directed planning in the training of autonomous agents is called model-based RL (Botvinick & Weinstein, 2019). We present such a model-based approach for planning by generating predictions based on a model of the world. This model can be used to simulate executing an action and thus to predict observation and reward for all possible actions. Based on the current observation and the predicted observations and rewards, the agent executes the next action. We conducted a model comparison between a model-free RL approach and a model-based RL approach, in which the agent is equipped with a model of the environment. In particular, we let both approaches compete against each other in a simulation environment (moonlander). In this environment, an agent has to successfully learn to steer a spaceship through a world while avoiding walls and randomly placed obstacles. In addition, in order to increase the unpredictability and difficulty of the task, there occasionally is drift pushing the spaceship to the left or right, respectively. We challenged the model-free and model-based RL agent in environments of different lengths and of three different difficulties. With increasing difficulty the amount of obstacles as well as the number of steps in which drift is applied increases. We assessed the performance in a final selection of 1000 levels per length-difficulty combination that have not been encountered before. Grouped by level length and difficulty, we measured performance in terms of how many steps the agents were able to take without crashing and the number of completely solved environments. In all applied configurations of level length and difficulty, the model-based approach outperformed the model-free RL agent in terms of the mean amount of steps the agent is able to take without crashing and the number of completed levels. The performance gap between the two agents became greater with increasing length and number of obstacles in environments. We conclude that goal-directed planning implemented by model-based RL enhances performance of an autonomous intelligent agent. We expect that combining both approaches in a hierarchical structure might yield a powerful architecture for acting in dynamic environments. A higher-level model-based RL agent that predicts action goals several time steps into the future passes these to a lower-level model-free RL agent executing them in the environment. This hierarchical structure reflects both forms of goal-directed and habitual action control present in humans (Dezfouli & Balleine, 2013).