Abstracts Track 2025

Area 1 - Risks and Ethical Concerns
Area 2 - Technology
Area 3 - Theory and Methods
Area 4 - Models
Area 5 - Applications
Nr: 154
Title:

Solutions of Equations and Systems of Equations with Various Omega Algebras
Authors:

Andreja Tepavcevic
Abstract: Omega-algebras are algebraic structures enriched with a generalized, lattice-valued equality, extending the classical notion of equality, introduced in [5]. In other words, Omega-algebra is an algebraic system containing an algebraic structure endowed with an Omega-equality, enabling the interpretation of classical identities as lattice-valued formulas. Quotients defined via level cuts of equality yield classical algebras that satisfy the same identities in the traditional sense. In recent years, various types of Omega-algebraic structures have been developed and investigated [1-4,6]. The distinction between classical fuzzy algebras and the broader, more flexible framework of Omega-algebras is that in classical fuzzy algebras, the construction begins with a set endowed with a fixed algebraic structure, like a classical group, and fuzziness is introduced through membership functions (obtaining, e.g., fuzzy groups, or subgroups). In contrast, in Omega-algebras we do not assume a predefined algebraic structure. While operations exist, they need not satisfy classical algebraic laws, allowing for a more adaptable modeling approach—particularly useful in contexts involving uncertainty or imprecise data. A key feature of this framework is the equivalence relation E connected to Omega-sets, which acts as a congruence in the algebraic context. When elements are indistinguishable within a certain p-cut, E identifies them as equivalent, facilitating a tolerant interpretation of data. This capability is essential in applications where exact precision must yield to the realities of incomplete or noisy information. In this context, we introduced and applied equations and systems of equations in various algebraic structures. While a quasigroup is the most general algebraic structure allowing unique solving of equations with one binary operation, fields are such structures with equations with two binary operations. Vector spaces are preferable structures for solving systems of linear equations. We developed the theory of solving equations in a lattice-valued setting with Omega-quasigroups, Omega-fields, and Omega-vector spaces [1-4]. In this lecture, the results obtained in different articles and in different teams will be presented and generalized in a unified and uniform manner, allowing for approximate solutions of various types of equations and systems of equations in this context. References. 1. Patricia Ferrero, Jorge Jiménez, María Luisa Serrano, Branimir Šešelja and Andreja Tepavčević. Omega vector spaces, submitted. 2. Jorge Jimenez, María Luisa Serrano, Branimir Šešelja, and Andreja Tepavčević. Omega ideals in omega rings and systems of linear equations over omega fields. Axioms, 12(8):757, 2023. 3. Jorge Jimenez, María Luisa Serrano, Branimir Šešelja, and Andreja Tepavčević. Omega-rings. Fuzzy Sets and Systems, 455:183–197, 2023. 4. Aleksandar Krapež, Branimir Šešelja, and Andreja Tepavčević. Solving linear equations by fuzzy quasigroups techniques. Information Sciences, 491:179–189, 2019. 5. Branimir Šešelja, and Andreja Tepavčević. Fuzzy identities. In 2009 IEEE International Conference on Fuzzy Systems, pages 1660–1664, 2009. 6. Branimir Šešelja and Andreja Tepavčević. Omega-groups in the language of omega-groupoids. Fuzzy Sets and Systems, 397:152–167, 2020

Area 1 - Risks and Ethical Concerns

Nr: 15
Title:

Agentic AI for Tourist Retrieval-Augmented Generation Recommendation
Authors:

Somatat Na Takuatung and Chokeanand Bussracumpakorn
Abstract: This paper explores the application of agentic Artificial Intelligence (AI) and Retrieval-Augmented Generation (RAG) in enhancing tourist recommendation systems, specifically for small-to-medium hotels. Traditional recommendation systems often suffer from data sparsity and the cold start problem. This research leverages Large Language Models (LLMs) and RAG to address these limitations to create a more robust and personalized recommendation system. The study introduces a shared business value model to foster a symbiotic relationship between hotels and their local communities, ultimately providing tourists with personalized and adaptive experiences. This model is implemented through a suite of applications, including the Hotel Front App, the Mobile Guest App, and the Community App, which facilitate data collection, service management, and community integration. Data were collected from a hotel in Koh Lanta, Krabi, to evaluate the system's performance, involving 50 participants. The system's effectiveness was assessed using key metrics such as Faithfulness, Answer Relevancy, and Maliciousness. The findings of this research demonstrate the significant potential of agentic AI and RAG to improve tourist recommendations, enhance guest experiences, and promote sustainable tourism practices. The paper also discusses the challenges and future directions of implementing such systems within the evolving travel ecosystem.

Area 2 - Technology

Nr: 167
Title:

C-SHAP for Time Series: A Concept-Based Approach to High-Level Temporal Explanations
Authors:

Annemarie Jutte, Faizan Ahmed, Jeroen Linssen and Maurice van Keulen
Abstract: Purpose: Explainable (XAI) aims to increase the transparency of AI systems. In literature, many XAI techniques for time series provide explanations using point-based attribution scores. SHAP is one such algorithm that can be used to determine attribution scores. However, localised methods fail to capture the influence of high-level patterns on model reasoning. Concept-based XAI provides explanations in terms of high-level features. In this research, we present C-SHAP, an approach that relies on SHAP to provide attribution scores for concepts. Contributions: Our first contribution is the presentation of two approaches to concept-based XAI for time series. The first approach is fully model-agnostic and post-hoc. The second approach relies on a concept-informed model, where the model is directly trained on the concepts under inspection. Our second contribution is the implementation of C-SHAP using multiple decomposition algorithms (Discrete Wavelet Transform, Empirical Mode Decomposition, and a custom decomposition) to provide different explanations. Our third contribution is a custom decomposition including components selected for human-centred interpretability: bias, trend, scale, low frequency, change in variance, and high frequency. Our fourth contribution is the validation of C-SHAP in two domains, human activity recognition and predictive maintenance, to showcase its generalizability. Methods: In our implementation, we construct concepts using time series decomposition. To determine the attribution of concepts, we apply SHAP using concept masking, where we determine the attribution of concepts by replacing them with uninformative concepts. C-SHAP is evaluated on three tasks: the classification of locomotion activities from accelerometer data in the OPPORTUNITY dataset, the detection of anomalous samples from engine noise in the FordA dataset, and the prediction of the remaining useful life of an engine from pressure levels in the PHM Turbofan dataset. We present anecdotal evidence and perform stability and feasibility analyses. Results: Using our custom decomposition, we find that `Bias' and `Scale' are specifically influential for locomotion models trained on the OPPORTUNITY dataset, `High frequency' is influential for anomaly detection models trained on the FordA dataset, and `Trend' is influential for the RUL prediction models trained on the Turbofan dataset. We argue these explanations match intuition and provide further validation through a feasibility analysis. Additionally, we provide an indication of stability for C-SHAP. Conclusion: In this research, we present a SHAP-based approach to concept-based XAI. C-SHAP is generally applicable to time series use cases, only requiring the selection of a concept construction algorithm. In future work, the selection of concepts for different domains should be explored and verified through user studies. Advantages of C-SHAP include its high-level explanations, the custom control over concept selection, and its post-hoc, optionally model-agnostic, explanations.

Area 3 - Theory and Methods

Nr: 108
Title:

Swarm-Based Training for Online Hyperparameter Optimization
Authors:

Jorge Alberto Calvillo, Katya Rodriguez and Carlos Ignacio Hernandez-Castellanos
Abstract: We introduce Swarm-Based Training (SBT), a novel population-based framework for online hyperparameter optimization in Reinforcement Learning (RL). Building upon the Population-Based Training (PBT) paradigm, SBT integrates Particle Swarm Optimization (PSO) dynamics to enhance exploration and adaptability. In this architecture, agents, which are modeled as velocity-driven particles, refine their hyperparameters asynchronously through local interactions and momentum-based updates. To address the limitations of greedy selection in standard PBT, we propose Swarm-Based Learning (SBL), a decentralized mechanism that enables underperforming agents to learn from their peers via pairwise performance comparisons. This bio-inspired approach yields emergent coordination, robust adaptation, and improved generalization without requiring global ranking or synchronization. We evaluated the SBT across six continuous control environments from the OpenAI Gym benchmark, using Proximal Policy Optimization (PPO) as the base algorithm. The results show that SBT outperforms state-of-the-art methods, including PBT, Population-Based Bandits (PB2), Generalized Population-Based Training with Pairwise Learning (GPBT-PL), and Random Search (RS), in terms of testing performance. Statistical significance was validated via median testing rewards and Critical Difference diagrams over 500 seeds, confirming the consistent superiority of SBT. The flexibility and modularity of SBT make it particularly well-suited for non-stationary, high-dimensional RL tasks. Asynchronous execution, decentralized adaptation, and principled swarm dynamics position SBT as a compelling advancement in AutoRL research.

Area 4 - Models

Nr: 170
Title:

AI Is the New UI: Introducing Agents to Accounting Software Products
Authors:

Ariane Goudie, Mahbub Gani and Nicholas Timon
Abstract: Recent advances in agentic frameworks are making possible a new kind of autonomous software that promises to revolutionise user experiences of software products. While significant progress is being made in AI technical infrastructure, little thought has been given to the interaction between the end-user and the agent; the needs of day-to-day operators with no prompting expertise are largely unmet. We propose an AI-first software approach in which agent-proposed actions are surfaced via a specialised Agentic interface, allowing the casual accounting software user to understand and confirm actions with ease. Accounting software for small to medium sized businesses tends towards extreme complexity, with users often requiring training to learn how to set-up and navigate unintuitive UI. Proficiency in the use of more advanced accounting features generally requires training spanning several months, which in turn is a significant cost to the organisation (https://taxcareacademy.co.uk/how-long-does-it-take-to-learn-bookkeeping-in-quickbooks/). The array of AI automation tools available in the accounting software market, ranging from automated data entry to transaction categorisation, affords the user only modest time savings from tedious accounting processes. Similarly, the recent wave of generative AI features in the form of chatbots does little to unlock the insights businesses need to transform their business operations; the cognitive burden is on the user to ask ‘good’ questions from an infinite pool of possible queries. Such a framework also presents software-builders with the immense technical challenge of building services capable of understanding the infinite variety of likely questions. The rise of the Agentic framework, however, is an opportunity to revolutionise the way users interact with their accounting software. The assigning of a specialised agent to each major accounting workflow is a paradigm-shift in which software begins to drive the user, replacing the universal application convention in which the user is in control and the software passively responds to their instructions. Working in tandem with regular ML tools and models, our proposed Agentic system will draw on their knowledge of the accounting domain and software APIs to orchestrate accounting workflows independently of the user, only prompting users for their input when complex scenarios arise. Harnessing its visibility of activity throughout the product, agents can prioritise tasks, proactively resolve discrepancies and anticipate blockers. The anticipated value to our business is twofold: first, the minimisation of user intervention; and second, meaningful optimisation of business operations beyond run-of-the-mill automation. Given recent predictions that up to 40% of Agentic projects will be shelved by 2027 ( https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027) it is becoming clear that we cannot ignore the UI, and and that generative AI indeed must be accompanied by a concomitant generative, dynamic UI for truly autonomous software to become a reality. The accompanying video to this abstract provides a concrete example of our AI-first UI implementation, demonstrating how business critical tasks are surfaced alongside one-click action controls, audit trails and a context panel revealing the intentionality for key Agentic decisions.

Area 5 - Applications

Nr: 37
Title:

Using Mutual Information and SHAP Analysis for Identifying Relevant Descriptors of the Synthesis of Perovskite Solar Cells
Authors:

Franklin Alexander Sepúlveda, Mónica Andrea Botero-Londoño and Byron Medina-Delgado
Abstract: Perovskite solar cells (PSCs) are emerging as a leading technology that may revolutionize the future of photovoltaics, thanks to their high power conversion efficiencies and low fabrication costs. However, PSC performance is highly sensitive to the synthesis conditions of the perovskite material. Information about the synthesis process and material structure is coded into variables known as descriptors or features. These descriptors enable the identification of the most influential synthesis conditions in terms of their impact on device performance. We remark that the traditional way to develop new materials is usually based on trial and error, which is resource-intensive and time-consuming; thus, knowing which variables are the most relevant can guide experimental efforts more efficiently. SHAP analysis has been used (scarcely) to quantify the contribution of synthesis descriptors to the performance of perovskite solar cells (PSCs) by interpreting machine learning (ML) models such as XGBoost. Although SHAP is a powerful method, it requires a trained model to compute feature importance, making the analysis dependent on the choice of ML model. In contrast, Mutual Information (MI) is a statistical measure that quantifies the amount of information shared between two variables. It is fully model-independent, without needing any predictive model. Additionally, we use Conditional Mutual Information (CMI), denoted as I(X; Y | Z), which measures the information gained about Y from X, given that Z is already known. These measures were estimated using the dataset reported in https://www.perovskitedatabase.com/ . Regarding the results, we present a comparative analysis of normalized global SHAP values and Mutual Information (MI) scores for several descriptors. Notably, both methods identify the DMF_DMSO_ratio as the most influential descriptor, indicating strong agreement on its predictive importance. Similar agreement is observed for features such as Perovskite_thickness and thermal_exposure, which also receive high scores in both analyses. However, discrepancies arise for descriptors representing the perovskite material (LLE_1 to LLE_4), which exhibit high MI values but low SHAP contributions. This suggests that, while these features carry significant information content, they may not be effectively utilized by the XGBoost model for prediction. In addition, we rank perovskite solar cell descriptors using Conditional Mutual Information (CMI). It places even greater emphasis on Perovskite_thickness and annealing-related parameters, reaffirming their central role in determining device performance. Interestingly, although the LLE descriptors show high MI values in Figure 1, their relatively low CMI scores suggest that much of their apparent relevance may be redundant once other variables are taken into account. This highlights how CMI refines feature importance by isolating non-redundant, conditionally informative descriptors, thereby improving the identification of truly impactful variables for modeling and interpretation in perovskite solar cell fabrication. Descriptors : DMF_DMSO_ratio: DMF and DMSO are the two most widely used solvents for the perovskite deposition procedure. annealing_XX: they encompasses the thermal treatment applied to the perovskite layer. LLE: Local Linear Embeddings for representing the perovskite materials. Perovskite_band_gap: energy difference between the material's valence band and conduction band.