Abstracts Track 2025

Area 1 - Artificial Intelligence
Area 2 - Agents

Area 1 - Artificial Intelligence

Nr: 120
Title:

Genetic Programming to Implement Glass-Box Models of Business Insolvency Prediction
Authors:

Ángel Beade, José Santos and Manuel Rodríguez
Abstract: The main objective of Bankruptcy Prediction Models (BPMs) is to anticipate the difficulties of a company and to provide a useful tool for decision-making by the multiple agents involved and the company itself. Bankruptcy Prediction is treated as a classification problem based on the information available in the financial statements of companies. In the last two decades, Machine Learning (ML) models have emerged as a better alternative to traditional models based on statistical models. Within the Evolutionary Computation methods in ML, we have been experimenting in recent years with Genetic Programming (GP) to implement BPMs for medium-sized companies. GP allows obtaining BPMs with high performance and stability over time (outside the learning period) [3], with comparable or better performance results with respect to other ML methods when used in conjunction with a Feature Selection (FS) process also based on GP [4]. Two complementary approaches of interpretability were pointed out by Miller [5]: i) explainability or interpretability, i.e., how well a human could understand the decisions of an ML model and ii) explanation, explicitly explaining decisions to people. In our application, the first approach has been considered with the choice of the tool itself (GP). GP provides straightforward interpretability of the optimized tree/program (at least human-readable solutions). The possibility to reduce the dimensionality of the input space by a Feature Selection process of the most relevant features also helps to improve the readability of the final selected BPMs. We have proposed different FS methods [1][4] with this goal. The second aspect was considered in [2] using multi-period BPMs, which predict failure over a range of future years (instead of failure at a specific future point in time), with the aim of interpretability for decision-making that helps to understand what actions need to be taken to counteract the failure of a company. Different scenarios are considered, with variations in financial magnitudes, and the marginal impact of each scenario is analyzed. This interpretability aspect in the BPMs, oriented to decision-making, is particularly useful when it can explain and inform future decisions aimed at counteracting the prediction of the BPMs. Therefore, these aspects allow a nuance to be introduced into the “black box” label of the proposed models. ACKNOWLEDGEMENTS Study funded by the Spanish Ministry of Science and Innovation (projects PID2020-116201GB-I00 and PID2023-148531NB-I00) and Xunta de Galicia (grant GPC ED431B 2022/33). REFERENCES [1] Beade, Á., Rodríguez, M. & Santos, J. (2023). Evolutionary feature selection approaches for insolvency business prediction with genetic programming. Natural Computing 22. [2] Beade, Á., Rodríguez, M. & Santos, J. (2023). Multiperiod bankruptcy prediction models with interpretable single models. Computational Economics. [3] Beade, Á., Rodríguez, M. & Santos, J. (2024). Business failure prediction models with high and stable predictive power over time using genetic programming. Operational Research 24:52. [4] Beade, Á., Rodríguez, M. & Santos, J. (2024). Variable selection in the prediction of business failure using genetic programming. Knowledge-Based Systems 289. [5] Miller, T. (2019). Explanation in AI: Insights from the social sciences. Artificial Intelligence 267.
Nr: 497
Title:

HikingTTE: A Deep Learning Approach for Hiking Travel Time Estimation Based on Personal Walking Ability
Authors:

Mizuho Asako, Yasuyuki Tahara, Akihiko Ohsuga and Yuichi Sei
Abstract: Hiking is a popular activity, but the risk of mountain accidents remains a critical issue. Accurately estimating hiking travel time is essential for enhancing safety and preventing such incidents. Existing methods for predicting hiking travel times have been based on modeling the relationship between the uphill and downhill slopes and walking speed. Recently, data-driven approaches have been proposed to model the relationship between slope and walking speed using GPS logs recorded by hikers. However, in existing hiking travel time estimation models, it is challenging to accurately reflect individual hiking ability and the effects of fatigue in travel time estimation. Therefore, this study proposes a travel time estimation method using hiking GPS logs that considers complex factors such as hikers' walking abilities, changes in walking pace, terrain, and elevation. Specifically, we applied deep learning-based travel time estimation techniques developed in the transportation field to hiking and built a model named "HikingTTE" that predicts hiking travel time based on hikers' GPS logs. The key idea of "HikingTTE" is to address the challenge of individual differences in walking speed that are unique to hiking. To achieve this, we integrated a slope-speed function based on a modified Lorentz function into a deep learning model consisting of LSTM and attention modules, so that the hiker's walking ability is reflected in the prediction. The task addressed in this study focuses on estimating the time required to travel from the current location to the goal while accounting for the walking performance from the start point to the current location, to adapt to changing conditions during hiking. The GPS log is divided into the first X% and the latter (100-X)%. The hiker’s walking ability is extracted from the walking performance in the first X% using a "slope-speed function based on a modified Lorentz function," the time required to complete the remaining (100-X)% is then predicted. In the experiments, we compared HikingTTE against existing hiking travel time estimation models and models adapted for this task using deep learning methods developed in the transportation field. As a result, HikingTTE reduced MAPE by 12.95 percentage points compared to the most accurate existing hiking travel time estimation method. Furthermore, compared to models developed in the transportation field, HikingTTE demonstrated an average reduction in MAPE of 0.97 percentage points. The HikingTTE achieved state-of-the-art performance in the hiking travel time estimation task. To the best of our knowledge, HikingTTE is the first attempt at travel time estimation incorporating a mechanism to account for individual walking ability. We anticipate that HikingTTE will set a new standard in hiking travel time estimation. One future challenge is to incorporate hikers' past hiking logs into the model. The code for the HikingTTE is available on GitHub at the following link: https://github.com/tarutaru2048/HikingTTE. See published papers for more information: https://cit.iict.bas.bg/CIT-2024/v-24-4/10341-Volume24_Issue_4-01_paper.pdf.
Nr: 15
Title:

How to Generate High-Dimensional Synthetic Medical Images with Little Data
Authors:

Michela Azzarito and Peter Krusche
Abstract: Brain Magnetic Resonance Images (MRI) are commonly collected in clinical trials for accurate diagnosis, effective treatment planning, and for monitoring disease progression in neurological diseases. Their analysis typically relies on human interpretation through expert readers such as radiologists and medical experts. Acquisition of MRI data is costly, time-consuming, and access to MRI data is often constrained by privacy concerns. As a result, datasets available for research and training machine learning models tend to be small, posing a major impediment to the development of reliable and generalizable AI algorithms. To mitigate these limitations, methods such as Generative Adversarial Networks (GAN) models have been proposed to generate synthetic medical images that can aid method development, data sharing and reproducibility while not being directly attributable to study subjects or individuals. In the context of synthetic brain MRIs, there are two main methodological challenges to overcome: (a)3-dimensional medical data are large objects; therefore, memory efficient methods are required, (b)limited amounts of data make it difficult to train neural networks. We addressed these challenges by using memory efficient method such as the HAGAN and adapted it to effectively manage small datasets by incorporating a Wasserstein gradient penalty loss function, resulting in the development of the WGP-HAGAN approach. The HAGAN model includes a hierarchical structure that simultaneously generates a low-resolution version of the full MRI and a randomly selected high-resolution sub-volume. The model includes 3 generators: a common generator (Ga), a generator for the low resolution (GL) and a generator for the high resolution (GH). In our experiments, we found that we needed to improve the training process in order to train on small datasets successfully. We explored different training strategies where we pre-train Ga and GL independently before training GH. In order to evaluate and compare strategies, we used quantitative metrics like precision, recall and Fréchet Inception Distance, computed using both low- and high-resolution synthetic data. In addition, we performed a visual inspection of the generated images and their diversity. Our experiments showed that using WGP-HAGAN with our progressive training process (first training low, then the high resolution generators) was the most stable out of the options we tested, with highest visual quality and recall achieved. We also find that training progressively allows us to focus learning of the diverse features in a population of brain images onto different parts of the model – e.g. representing diversity in the low or high resolution parts exclusively. To enhance transparency and encourage collaboration and re-use within the research community, we offer a publicly accessible code repository. This repository includes comprehensive scripts for data preprocessing, GAN training, and evaluation of both low- and high-resolution synthetic brain MRI. Alongside these scripts, we provide pre-trained models and detailed documentation to facilitate the reproduction of our results and the adaptation of our methodologies by other researchers. Additionally, we present an end-to-end tutorial covering MRI generation using these trained models. This tutorial is designed to assist in data augmentation for various downstream applications, including potentially applying to other 3D medical imaging modalities.
Nr: 449
Title:

Federated Learning Algorithm to Suppress Occurrence of Low-Accuracy Devices
Authors:

Koudai Sakaida, Keiichiro Oishi, Yasuyuki Tahara, Akihiko Ohsuga and Yuichi Sei
Abstract: In Federated Learning (FL), collaborative model training occurs across multiple devices while preserving data privacy. Typically, a central server randomly selects a subset of devices in each training round, provides them with the current model, and each selected device updates it using its own local dataset. The models are then sent back to the server and aggregated, forming an improved collaborative model as learning progresses. However, under Non-IID data conditions, this process can become imbalanced, leading to noticeable variations in model performance across devices. In such scenarios, some devices suffer from significantly lower accuracy, thereby undermining performance fairness. Addressing underperforming devices is therefore essential for maintaining fairness and ensuring the overall effectiveness of FL systems. This research proposes a novel approach to suppress low-accuracy devices in FL by integrating reinforcement learning, specifically employing a Double-DQN for device selection. The method introduces a reward mechanism based on the Euclidean distance between each device's local model and the global model, thereby encouraging the selection of devices that positively influence global performance. To manage high-dimensional model weights and reduce computational complexity, UMAP is utilized for dimensionality reduction, compressing weights into a lower-dimensional space while preserving important information. The approach also adopts multi-action selection, allowing the simultaneous selection of multiple devices in each communication round, thereby improving learning efficiency. Experiments were conducted on datasets including CIFAR-10, GTSRB, and synthetic datasets. Non-IID conditions were simulated for CIFAR-10 and GTSRB using Dirichlet distributions, while the synthetic datasets were generated following methods inspired by prior studies, controlling model and data heterogeneity by adjusting parameters. All the datasets used are standard benchmarks for evaluating the method's effectiveness. Results show that the proposed method improves the average accuracy of the bottom 10% of devices by up to approximately 4% without reducing the overall average accuracy compared to existing methods including FedAVG, q-FFL, FedFE, and FedHEAL. In highly Non-IID environments, the improvement in low-accuracy devices is particularly significant. Analysis of device selection patterns indicates that the reinforcement learning agent increases the selection frequency of devices with more diverse and abundant data, contributing to performance improvement of underperforming devices. Generally, given that only a subset of devices is selected in each round of FL, devices with more diverse data tend to be preferentially chosen. Future work includes extending the reinforcement learning approach to not only device selection but also weighted model aggregation by employing multi-agent reinforcement learning. This extension aims to further suppress the occurrence of low-accuracy devices. Additionally, addressing practical considerations such as computational and communication costs, and privacy concerns is essential. Developing new algorithms that effectively suppress low-accuracy devices in Non-IID environments while considering these constraints remains a critical area for future research. The full version of this work has been submitted to the International Journal of Electrical and Computer Engineering Systems.
Nr: 458
Title:

CLEaR: Counterfactuals Through Large Language Models for Enhanced Recourse
Authors:

Yash Narendra and Niladri Chatterjee
Abstract: In the realm of machine learning explainability, traditional approaches often provide universal, one-size-fits-all explanations that may not effectively resonate with individuals facing model decisions. This abstract introduces CLEaR (Counterfactuals through Large Language Models for Enhanced Recourse), proposing a paradigm shift towards dynamic and tailor-made explanations, challenging the static nature of current explanatory frameworks. We advocate for a more personalized and adaptive approach, emphasizing the need to cater to the unique contexts and preferences of individuals. The central idea revolves around the belief that explanations in machine learning should evolve dynamically, specifically tailored to the individual’s needs. Counterfactual explanations, as a means of enhancing interpretability, should not adhere to a static template but instead address the nuanced concerns of each user. Highlighting the limitations of static explanations, this work proposes a departure from conventional practices by integrating dynamic personalization in the generation of counterfactual explanations. By leveraging Large Language Models (LLMs), we present a methodology that allows individuals to actively participate in the creation of counterfactuals, tailoring explanations to their unique preferences and contexts. We introduce an interactive user interface that facilitates the generation of dynamic, personalized counterfactual explanations. Through empirical analysis, we demonstrate the feasibility and efficacy of this approach, emphasizing its transformative potential in enhancing user understanding and trust in machine learning models. This abstract asserts that explanations should no longer be static and uniform but should embrace dynamism, catering to the diverse needs of individual users. By championing the cause of dynamic, personalized explanations, we anticipate a significant step forward in making machine learning models more transparent, interpretable, and user-centric.
Nr: 495
Title:

Polyphenol-Specific CPI Prediction Using GNNs with LLM Embeddings and Transfer Learning
Authors:

Shota Majima, Koyo Fujisaki, Kengo Morohashi and Munehiro Takimoto
Abstract: In drug discovery, it is essential to identify compounds that can bind to target proteins and modulate their function. However, it is costly to experimentally test the interactions between a large number of protein–compound pairs. Virtual screening addresses this challenge by computationally predicting protein–compound interactions, enabling large-scale, low-cost, and rapid identification of promising lead compounds. Polyphenols, which consist of aromatic rings and/or hydroxyl groups and are mainly synthesized in plants, are particularly noteworthy due to their potential therapeutic benefits, including a reduced risk of cardiovascular disease and cancer. Their antioxidant properties and capability to regulate cellular signaling pathways have made them a focus of research in functional foods and disease prevention. Despite these promising features, the molecular mechanisms by which polyphenols interact with proteins remain poorly understood, largely due to the limitations of general compound-protein interaction (CPI) prediction models in capturing the unique chemical and biological characteristics of polyphenols. This gap gives the need for prediction models tailored to polyphenols. In this study, we propose a novel approach using Graph Neural Networks (GNNs) and transfer learning for developing a polyphenol-specific CPI prediction model. The GNN architecture is designed to effectively capture the structural and relational properties of compound-protein networks, which makes it well-suited for modeling complex interactions. Our method integrates embeddings derived from pre-trained large language models (LLMs), which are optimized for representing both chemical and biological features, to the GNN. The integrated model was first pre-trained on general compounds using data from the STITCH database, which includes compound-compound and compound-protein interaction networks, and the STRING database, which provides protein-protein interaction networks. This pre-training phase allowed the model to learn generalizable features of compound-protein interactions. Subsequently, the model was fine-tuned specifically for polyphenol-related tasks using transfer learning. We evaluated the model’s performance on a polyphenol-specific task involving previously unseen compounds. This scenario assessed the model's capability to predict interactions between new polyphenol compounds and proteins. We used AUROC and AUPRC as evaluation metrics for the evaluation, of which the results demonstrated that the transfer learning-based model outperformed the non-transfer-learning model. This shows the effectiveness of incorporating polyphenol-specific features into the model through transfer learning. The relationship between polyphenols and biological activities suggested by this model may contribute to understanding of their therapeutic roles. Furthermore, this model could inform the design of functional foods and therapeutic agents, which contribute to advancements in medicine. Additionally, this approach can be applied to develop prediction models for other specific compound groups. Future work will focus on integrating additional data sources (e.g., gene ontology-based networks) and validating predictions through in vitro and in vivo experiments. It enables us to further refine the model, move toward clinical applications, and ultimately unlock the therapeutic potential of polyphenols while establishing a new framework for compound-specific CPI prediction.
Nr: 496
Title:

AdaSniper: Toward Automated Detection of Systematic Errors in Image Recognition in Open World by Adaptive Exploration with Generative AI
Authors:

Yuma Torikoshi, Fuyuki Ishikawa, Yasuyuki Tahara, Akihiko Ohsuga, Yuichi Sei, Juichi Takahashi and Yohei Takagi
Abstract: Image recognition systems (IRS) deployed in open-world environments often exhibit systematic errors, i.e., recurring mistakes triggered by specific situational or contextual factors rather than random noise. For example, it has been reported that a certain company’s autopilot system in automobiles tends to misrecognize scenes at night when emergency vehicles, such as police cars, place traffic cones to regulate traffic. Such an error may lead to accidents, underscoring the importance of detecting and addressing systematic errors before deployment in real-world scenarios. Traditional testing approaches, including metamorphic testing or attribute-based evaluations, rely heavily on predefined ontologies and perturbation operators. While these approaches can uncover some vulnerabilities, they fail to detect errors emerging from unexpected conditions due to their limited scope. With the advent of generative AI (GenAI), including Large Language Models (LLMs) and text-to-image (T2I) models, testers can now flexibly generate a wide range of semantic test scenarios and corresponding images, free from rigid, predefined categories. In this paper, we present AdaSniper, an adaptive, fully automated testing framework that leverages both LLMs and T2I models to systematically identify and magnify the most error-prone conditions of an IRS. AdaSniper’s closed-loop process initiates test scenarios via an LLM, creates corresponding images with a T2I model, and runs the IRS on these generated inputs. It then analyzes failure patterns—such as frequent misclassifications or missed detections—and strategically feeds back detailed misclassification data (including the error destination classes) to the LLM. Guided by this feedback, the LLM proposes increasingly challenging scenarios that sharply intensify the IRS’s weaknesses. Crucially, AdaSniper accomplishes this iterative refinement without human intervention, rapidly homing in on high-failure conditions that conventional, more rigid testing methods can easily overlook. We conducted evaluations on both classification and object detection models trained on large-scale datasets. Compared to a baseline method that does not exploit such detailed misclassification data, AdaSniper consistently and substantially outperformed it in detecting high-error-rate scenarios. In practical terms, AdaSniper discovered harder-to-detect errors significantly faster and maintained a growing edge as testing progressed. Beyond merely driving up error detection rates, we experimented with strategies that also ensure diversity in discovered scenarios. By adjusting the prompts and reference data provided to the LLM, AdaSniper can be tuned to find a broader spectrum of test conditions, enabling a more well-rounded evaluation. These findings highlight that LLM-driven scenario generation, when coupled with T2I-based input synthesis and adaptive feedback loops, can form a new paradigm in open-world IRS testing—one that not only unearths systematic errors more efficiently but also flexibly adapts to various testing objectives. This methodology shows promise not only for image classification and detection tasks, but also for other domains such as autonomous driving, medical imaging, and out-of-distribution detection, laying the groundwork for more robust and trustworthy AI systems.
Nr: 506
Title:

AI and the Education of Gen Z: Transforming Tools into Companions
Authors:

Marc Cleiren, Julian van der Kraats, Jaap van den Herik and Peter van der Putten
Abstract: -The upgrade of Gen Z- Our multidisciplinary team investigated how to empower current and future generations in working with AI as companion rather than seeing it merely as a production tool. The research focused on making AI integration beneficial for educators and students by developing new skills in their academic and professional activities. Generation Z (born 1997-2012) are already digital natives, and now transfer into "Generation ZAI" by incorporating AI as an extension of their cognitive processes during their formative years in higher education. As one student put it: “I can no longer imagine how I ever studied without AI”. This transition challenges traditional academic skills and knowledge acquisition methods, with students becoming knowledge managers rather than knowledge carriers. -Our research question- How can educators and students profit from embracing AI as a companion to support the current generation in (a) developing new skills and (b) successfully applying them in future jobs? -Methods- We combined insights from educational and neurofunctional psychology on how human motivational systems (inter)act. We used insights on Behavioral Approach Systems (BAS)/Behavioral Inhibition Systems (BIS) (1) and System 1 and 2 thinking (2). We then designed and conducted 3-hour multidisciplinary AI hackathons and a 3-month course (in 2023-2024). Students from all disciplines were tasked to design AI applications, not requiring programming knowledge, focusing foremost on creative and formative processes rather than outcomes. -What did we learn?- Key observations are that participants overcame initial hesitation through learning-by-doing approaches and iterative interactions with Generative AI. We identified four critical areas for development: transforming AI from a production tool to creative companion, empowering productive AI collaboration, preserving and encouraging human interaction skills, and maintaining problem-solving confidence while working with AI. First findings suggest that adjustments to academic teaching practices and assessment may be necessary, but also fruitful to support Generation ZAI's development, with recommendations for iterative improvements in AI-integrated educational approaches. Special attention should be given to help students transition and remain in a ‘learning’ than ‘production’ mode. -Conclusions and presentation- In our presentation we discuss the data and address: a) Shifting Gen ZAI’s use of AI as production tool towards more creative use. b) Empowering students and staff to productively collaborate with AI instead of only consuming AI results. c) Using and preserving human-to-human interaction skills as agentic in working with AI. d) AI enhanced academic learning without undermining self-confidence in problem solving and analytical skills Our approach currently serves as good practice at our University, serving as a model for AI formation of our staff and students. -References- (1) Beek, I., Kranenburg, I., Taris, T., & Schaufeli, W. (2013). BIS- and BAS-activation and study outcomes: A mediation study. Personality and Individual Differences, 55, 474-479. https://doi.org/10.1016/J.PAID.2013.04.013 (2) Schwab, K., Meerts-Brandsma, L., Lackey, Q., Povilaitis, V., Zajchowski, C., Moseley, B., & Dustin, D. (2020). Learning, Fast and Slow. SCHOLE: A Journal of Leisure Studies and Recreation Education, 37, 186 - 193. https://doi.org/10.1080/1937156X.2020.1806760.
Nr: 512
Title:

Time-Dynamic Cherry Blossom Prediction Using First-Order Reduced and Controlled Error Trained Echo State Networks
Authors:

Gantogoo Oyunbat, Erik Syniawa and Fred Hamker
Abstract: The human brain performs complex predictive tasks while consuming minimal energy, highlighting the vast efficiency gap between biological and artificial computing systems. While modern computers offer impressive computational capabilities, their energy requirements limit widespread deployment of predictive models. Neural systems inspire reservoir computing approaches that aim to bridge this gap, combining computational power with energy efficiency [1]. We demonstrate this potential by applying reservoir computing to a real-world time-dynamic forecasting challenge where traditional regression models struggle with inherent non-linearities. Accurate cherry blossom date prediction is crucial for cultural, ecological, and economic planning in Japan. This study uses an Echo State Network (ESN) with FORCE (First-Order Reduced and Controlled Error) [2] recursive least squares training to model the time-dynamic relationship between temperature, humidity, geographical coordinates, and cherry blossom dates. Public datasets, including cherry blossom records and meteorological data, train the ESN to predict bloom events in Japanese cities. The preprocessing pipeline included data cleaning, handling missing values, feature scaling, and aligning the data with the ESN’s temporal firing dynamics. The ESN processes temperature and humidity data incrementally while location data remains fixed, reflecting the progressive availability of environmental information. Results show that prediction accuracy consistently improves with longer input sequences for both first and full bloom predictions. FORCE training efficiently adapted the readout layer to the non-linear, dynamic input-output relationships. Even with limited training data, our ESN maintains stable performance and achieves accuracy comparable to transformer-based models implemented in the darts package for predicting cherry blossom dates across Japanese cities, while requiring significantly fewer learnable parameters. This approach demonstrates potential for energy-efficient hardware implementations as ESNs are particularly suitable for neuromorphic chips due to fixed internal weights and simplified training. The study validates reservoir computing as an efficient alternative for time-dynamic forecasting and points to scalable, energy-efficient hardware solutions for real-world problems, bridging the gap between biological and artificial neural computation. The code is available at https://github.com/esyniawa/rc-prediction_cherry_blossom. [1] M. Yan, C. Huang, P. Bienstman, P. Tino, W. Lin, and J. Sun, “Emerging opportunities and challenges for the future of reservoir computing,” Nature Communications, vol. 15, no. 1, p. 2056, 2024. [2] D. Sussillo and L. F. Abbott, “Generating coherent patterns of activity from chaotic neural networks," Neuron, vol. 63, no. 4, pp. 544-557, 2009.

Area 2 - Agents

Nr: 508
Title:

Incentive Policies with Deceptive Agents and a Learning Regulator
Authors:

Andrea Araldo, Amir Brudner, Mark Daychman and Ravi Seshadri
Abstract: We consider a regulator aiming to encourage sustainable behavior among agents, including individuals and businesses. Individuals face choices regarding goods, food, services, and mobility, while businesses make decisions on production modes, internal organization, and more. However, agents’ selfish utility-maximizing choices often conflict with sustainability. To promote sustainable choices, the regulator can use tailored "measures", such as incentives, subsidies, prices, taxes, or bans. Personalized policies adjust these measures to each agent’s needs and preferences [1]. To implement a personalized policy, the regulator must learn agents’ preferences by observing their past choices. Proposed personalized policies rely on the classic discrete choice modeling assumption that agents are rational and honest, always maximizing their utility. However, this assumption does not hold for personalized policies, where agents may behave deceptively to hide their true preferences and manipulate the regulator to obtain a “deception premium,” i.e., a more favorable measure than they deserve. Recent literature models deceptive behavior as follows: a deceptive user "selects a simulated utility" different from their true one and makes decisions to maximize it to mislead the regulator [2]. The standard assumption is that the regulator immediately knows the user’s simulated utility and provides measures accordingly. However, this assumption is overly pessimistic, as it makes misleading the regulator too easy. In this paper, we adopt a more realistic assumption that the regulator learns the user’s utility from their observed choices. Therefore, to mislead the regulator, the user must consistently act in line with the simulated utility, sacrificing the maximization of their true utility. This results in a cost of deception, i.e., the loss from not maximizing their true utility. It remains unclear when this cost is lower than the deception premium. Our overarching aim is to analyze and enhance the robustness of personalized incentive policies for deceptive agents. While regulators can prevent any deception by withholding incentives, this approach does not shift users' behavior toward sustainable choices and thus will not improve social welfare. On the other extreme, excessive incentives may encourage users' deceptive behavior. Hence, the question we raise is, "What is the optimal level of incentive that maximizes social welfare, along with the corresponding level of deception, if any?" To tackle this question, we formalize the setting in a game-theoretical framework as a leader-follower game. We model the regulator's process of learning users' preferences as a Markov Chain and users' sequential decision-making as a Markov Decision Process. This formulation allows us to characterize optimal user strategies, i.e., the sequence of choices and the level of incentive below which deceptive behavior is discouraged. These characterizations can inform the design of incentive policies that limit deceptive behavior to ensure social welfare improvement. References [1] Y. Xie, R. Seshadri, Y. Zhang, A. Akinepally, M. E. Ben-Akiva, Real-time personalized tolling for managed lanes, Transportation Research Part C: Emerging Technologies 163 (2024) 104629. [2] Q. Dawkins, M. Han, H. Xu, The limits of optimal pricing in the dark, Advances in Neural Information Processing Systems 34 (2021) 26649–26660.
Nr: 513
Title:

Nudge & Boost Agents: Bridging Physical and Metaverse Environments through Agent Technology to Effectively Influence Human Decision Making
Authors:

Tetsuo Ono
Abstract: Behavioral economics has revealed that human decision-making is not as rational and logical as previously assumed [Kahneman 2011]. Our research focuses on investigating the types of information technologies and insights that can be utilized to construct environments and contexts to enhance decision-making. Furthermore, we aim to facilitate behavioral adaptation in response to contextual factors. The environments and contexts discussed in this paper encompass both the physical world and metaverse spaces. Recent studies have demonstrated that metaverse experiences can significantly influence real-world behavior [Sakamoto 2024]. This paper adopts a broader perspective on decision-making environments and aims to design environments that facilitate effective human decision-making. We first present the design of an ambient intelligence system utilizing agents to support appropriate decision-making in the physical world [Ono 2023]. Specifically, we propose a methodology for designing "Nudge & Boost Agents" that integrate the behavioral economic concepts of "nudge" and "boost" to encourage behavioral adaptation to situations (nudge function) and develop individual decision-making skills (boost function). The Nudge & Boost Agents are implemented based on ambient intelligent systems grounded in previous Human-Agent Interaction research. This paper describes the proposed agent concept, system implementation, and demonstrates system effectiveness through experimental evaluation results. Furthermore, this paper elucidates decision-making processes involving agents in metaverse spaces and aims to develop systems that can support such processes [Chenling 2023]. We particularly focus on perspective-taking, which enables consideration of others' thoughts and objectives, and has been recognized as crucial for facilitating smooth human-human interaction. In the research presented [Hang 2023], we investigated how decision-making patterns change in VR environments when experimental participants adopt either an altruistic perspective (helping robots) or a recipient perspective (being helped by robots). Results indicated that participants who adopted an altruistic perspective maintained their altruistic viewpoint toward helping robots. Conversely, participants who received help from robots did not demonstrate similar altruistic decision-making. This study suggests that human-robot interaction experiences in VR environments influence subsequent perspective-taking and prosocial decision-making within VR contexts. By elucidating decision-making processes and their interactions in both physical and metaverse spaces, we can develop decision support systems that guide humans toward enhanced well-being.