ICAART_DC 2023 Abstracts

Full Papers
Full Papers
Paper Nr: 5
Title:

Incorporating Sense of Control in Dynamic Multi-Tasking Problems: A Model-based Hierarchical Reinforcement Learning Approach
Authors:

Annika Österdiekhoff
Abstract: Autonomous intelligent agents often fail in complex dynamic multi-tasking scenarios that are easily solvable for humans. For example, agents fail to adapt to new situations or cannot coordinate task switching between multiple tasks. A reason for this inability could be that in comparison to humans intelligent agents are missing cognitive capabilities. For one thing, humans have a sense of control (SoC), the feeling of being in control for successfully solving tasks. In general, it is unclear how to model the SoC of humans and how to make use of it for autonomous intelligent agents. This paper will present an approach to model the SoC and incorporating into autonomous intelligent agents modeled by model-based hierarchical reinforcement learning. Moreover, planned studies on the SoC in humans in multi-tasking problem scenarios are discussed. The hypothesis is that using this approach and equipping autonomous intelligent agents with a SoC will lead to improved decision-making in complex dynamic multi-tasking problems.
Paper Nr: 6
Title:

AI, Fiction and the A(i)uthor: What Could that Mean to Readers and Literary Theorists?
Authors:

Ines Bouteldj
Abstract: The use of artificial intelligence (AI) in the creation of literary texts (fiction and poetry) has raised important questions about the nature of authorship, reading and meaning-making. In this research, I aim to explore these questions in the context of literary studies and criticism with a particular focus on the impact of AI-generated novels on the field and on the concept of authorship in literary theory. I also investigate the impact of AI-generated texts on reading and meaning-making processes. Finally, I consider the possibility of AI achieving creativity and the potential challenge it poses to traditional notions of creativity being a solely human quality.
Paper Nr: 7
Title:

Defining and Analyzing Java Program for Automatic Singleton Design Pattern Injection
Authors:

Abir Nacef
Abstract: One promising approach in software engineering is to improve the quality of code and its design by relying on pre-established restructuring (refactoring). In particular, the injection of a design pattern is considered a form of complex refactoring. To be able to apply these restructurings, it is necessary to identify specific structures and behaviors in the source code. The goal of our approach is the definition of implicit implementation of singleton design patterns (SII). The definition of this implementation makes easy their detection, and then the injection of the pattern. By analyzing the singleton pattern (SP) intent, we define a set of information that can make SII. Then, for analyzing the source code and extracting this information, we apply a structural and semantic analysis with the LSTM model for text classification. The LSTM model is trained by a created dataset named DSII (Data for Singleton Implicit Implementation) specific to each information in the source code (features). For evaluating different LSTM models, we create and label a new data named SDE (Singleton Evaluation Data) collected from the GitHub Java project. The empirical result proves that our used LSTM model can successfully extract proposed feature values with better results compared to SVM and Naive Bayes techniques
Paper Nr: 8
Title:

Adaptive Market-Making Using Reinforcement Learning Within a Multi-Agent System
Authors:

Christopher Cho
Abstract: The financial markets are one of the most complex and well-studied multi-agent systems that exist in the world. For nearly half a century, researchers have carried out studies to learn more about the properties of the capital markets, which have become an essential part of the modern-day economy. However, despite the attention the field has received over the past decades, the task of creating a simulated market environment with realistic behaviour remains incomplete. This poses a critical problem when formulating new research hypotheses and developing new trading strategies. Actions taken in the market cannot always be assumed to be deterministic or ”small” in magnitude and in turn, have a negligible effect on the market dynamics. In this project, we first address the necessity for a simulated market environment by presenting a novel multi-agent system capable of generating a series of market actions representative of a real-world exchange. Then, we develop and compare market-making strategies that are controlled by Reinforcement Learning algorithms by deploying them into the multi-agent simulation to assess the performance of these market-makers in a fair, dynamic and realistic setting.
Paper Nr: 9
Title:

TeNet: on Converting Unstructured Text to Network for Understanding and Supporting Partnership in International Business Markets
Authors:

Didier Gohourou
Abstract: Network representation of data is key to a variety of fields and their applications including trading and business. A major source of data that can be used to build insightful networks is the abundant amount of unstructured text data available through the web. The efforts to turn unstructured text data into a network have spawned different research endeavors including the simplification of the process. This study presents the design and implementation of a pipeline that turns unstructured text data into a graph, aiming to minimize human intervention. It describes the application of natural language processing techniques used to process the text and learning algorithms that categorize the nodes and the links. The study also make use of graph learning techniques in recommendation mechanisms. The methodology is proposed as a blueprint for future studies that need a domain-specific network representation of unstructured text data, especially the ones trying to solve international business partnership challenges.
Paper Nr: 10
Title:

More Accessible Explainable Artificial Intelligence with Human-in-the-Loop Approach
Authors:

Maciej Mozolewski
Abstract: I present a Human-in-the-loop (HITL) approach to Explainable Artificial Intelligence (XAI), which is the subject of my planned PhD desideration. I argue that it allows for the incorporation of expert knowledge into Machine Learning-based Artificial Intelligence systems. In addition, HITL allows for greater engagement and accessibility of end-users, such as domain experts, eCommerce owners, etc. It could help XAI to be more widely adopted in real usage scenarios. I briefly present two of my own works that represent this approach. The first one deals with the use of XAI metrics such as Stability, Consistency and Perturbational Accuracy Loss using the Intelligible eXplainable AI framework (InXAI) package. In this work, I show how XAI metrics can improve an ensemble of classifiers. The second work is an example of a HITL prototype for clustering analysis carried out within the Knowledge Augmented Clustering (KnAC) project. It introduces an intuitive, yet under-explored in the literature, distinction between objective data and metadata and shows how it can help in creating discourse-based explanations.