Departamento de Ciencias de la Computación e Inteligencia Artificial

IPOP-CMA-ES and the Influence of Different Deviation Measures for Agent-Based Model Calibration

2026-04-08T06:47:23Z

IPOP-CMA-ES and the Influence of Different Deviation Measures for Agent-Based Model Calibration Vargas Pérez, Víctor Alejandro; Chica Serrano, Manuel; Cordón García, Óscar Calibration is a crucial task on building valid models before exploiting their results. This process consists of adjusting the model parameters in order to obtain the desired outputs. Automatic calibration can be performed by using an optimization algorithm and a fitness function, which involves a deviation measure to compare the time series coming from the model. In this paper, we apply a memetic IPOP-CMA-ES for the calibration of an agent-based model and we study the effect of different deviation measures in this calibration problem. Classical metrics calculate the mean point-to-point error, but we also propose using an extension of dynamic time warping, which considers trend series evolution. In order to determine if calibrating with an specific metric leads to better solutions, we carry out an exhaustive experimentation by including statistical tests, analysis on the values of the calibrated parameters, and qualitative results. Our results show IPOP-CMA-ES obtains better performance than a genetic algorithm. In addition, MAE, MAPE and Soft-DTW are the metrics which report best results, although we get a similar behavior for all of them. This work is supported by the Spanish Agencia Estatal de Investigación, the Andalusian Government, the University of Granada, European Regional Development Funds (ERDF) under grants EXASOCO (PGC2018-101216-B-I00), SIMARK (P18-TP-4475), and AIMAR (A-TIC-284-UGR18) as well as by the Program "Becas de Iniciación UGR - Banco Santander".

Social network of peer-to-peer accommodations for a visual decision support system in tourism: The case of the Canary Islands

2026-04-08T06:36:43Z

Social network of peer-to-peer accommodations for a visual decision support system in tourism: The case of the Canary Islands Vargas Pérez, Víctor Alejandro; Cordón García, Oscar; Chica Serrano, Manuel; Hernández Guerra, Juan María The peer-to-peer accommodation market has experienced significant growth in recent years, leading to increased competition and offer heterogeneity. This scenario presents challenges for investors and stakeholders, required to value the importance of differentiation and accommodations' typology to ensure favorable profits and social impact. In this work, we examine the touristic accommodation market in the Canary Islands using real data from Airbnb and applying a novel network-based visual methodology. The data analysis methodology involves the creation and visualization of a network that places accommodations based on their similarity. Using community detection algorithms, we identify accommodation typologies, perform a descriptive analysis of the resulting clusters, and evaluate economic and exogenous variables. Nine accommodation types are found having key differentiating characteristics such as guest capacity, number of properties owned by the host, and managerial aspects (for example, cancellation policy). Clusters with higher economic benefits (characterized by a large capacity) are placed on the periphery of the visual map in contrast to common accommodation types, located in the center; thus showing the importance of differentiation. The accommodations' typologies are not specific to a particular island, but are homogeneously distributed in the Canaries archipelago. The results emphasize the managerial advantage of this decision support system for investors and tourist managers in making informed strategic decisions. This work was supported by MCIN/AEI/10.13039/501100011033 and ERDF “A way of making Europe” under grant CONFIA (PID2021-122916NB-I00); and by Ministerio de Industria y Turismo of the Spanish Government under grant TUR-RETOS2022-075. V. Vargas is also supported through the FPU program (FPU20/02441). M. Chica is also supported by grant EMERGIA21_00139 funded by Consejería de Universidad, Investigación e Innovación of the Andalusian Government.

Opinion Dynamics with Highly Oscillating Opinions

2026-04-07T10:42:29Z

Opinion Dynamics with Highly Oscillating Opinions Vargas Pérez, Víctor Alejandro; Giráldez-Cru, Jesús; Cordón García, Óscar Opinion Dynamics (OD) models are a class of agent-based models that describe how opinions evolve within a population. In these models, individual opinions change through interactions governed by an opinion fusion rule that specifies how updates occur. Despite their simplicity, OD models offer interpretable mechanisms for understanding the collective dynamics of opinion formation. However, most existing approaches focus on the emergence of consensus, fragmentation, or polarization, while overlooking real-world scenarios characterized by highly oscillatory trends. This study addresses this limitation by evaluating the ability of several OD models extended with dynamic parameters to reproduce oscillatory dynamics. To this end, we formulate an optimization problem solved via evolutionary algorithms. The methodology is first validated on synthetic target series to assess the intrinsic oscillatory capabilities of the models, and subsequently applied to a real-world dataset of public opinion about immigration, drawn from the monthly barometer of the Spanish Sociological Research Center. Results show that the Agent-independent Time-based Bounded Confidence and Repulsion (ATBCR) model, which combines confidence-based and polarization-based update mechanisms, achieves the best performance. The optimized model closely reproduces historical opinion fluctuations while exhibiting interpretable, human-like patterns of collective evolution.

Torchmil: A PyTorch-based library for deep multiple instance learning

2026-03-23T12:36:33Z

Torchmil: A PyTorch-based library for deep multiple instance learning Castro Macías, Francisco M.; Sáez Maldonado, Francisco J.; Morales Álvarez, Pablo; Molina Soriano, Rafael Multiple Instance Learning (MIL) is a powerful framework for weakly supervised learning, particularly useful when fine-grained annotations are unavailable. Despite growing interest in deep MIL methods, the field lacks standardized tools for model development, evaluation, and comparison, which hinders reproducibility and accessibility. To address this, we present torchmil, an open-source Python library built on PyTorch. torchmil offers a unified, modular, and extensible framework, featuring basic building blocks for MIL models, a standardized data format, and a curated collection of benchmark datasets and models. The library includes comprehensive documentation and tutorials to support both practitioners and researchers. torchmil aims to accelerate progress in MIL and lower the entry barrier for new users.

Mitigating Linguistic Aggression in Group Decision-Making: A Comparative Analysis of AI-Driven Hostility Detection

2026-03-16T10:09:59Z

Mitigating Linguistic Aggression in Group Decision-Making: A Comparative Analysis of AI-Driven Hostility Detection Trillo Vílchez, José Ramón; González-Quesada, Juan Carlos; Cabrerizo Lorite, Francisco Javier; Pérez Gálvez, Ignacio Javier The process of group decision-making is an integral component not only for quotidian interactions but also for strategic deliberations. However, it is profoundly shaped by the inherent semantic indeterminacy of natural language. This linguistic ambiguity starkly contrasts the syntactic and semantic precision characteristic of machine-generated language. Furthermore, the conveyance of affective states–such as aggressiveness or elation–via natural language introduces a layer of complexity that can significantly perturb the equilibrium of the group decision-making process. In response to these challenges, we propose an advanced consensus-reaching methodology based on sentiment analysis to quantify and mitigate aggressiveness in discourse. This study conducts a comparative evaluation of three state-of-the-art large language models: Gemini, Copilot, and ChatGPT for their efficacy in detecting and assessing hostility. By calibrating the influence of individual participants based on their degree of linguistic aggression, the proposed framework attenuates the disproportionate impact of dominant voices, thus fostering a more balanced and equitable deliberative environment. This methodological innovation not only incentivizes the adoption of a more dispassionate and constructive linguistic register but also safeguards the integrity of collective decision-making processes against the distortive effects of undue emotional influence. Across five repeated evaluations per comment, ChatGPT and Gemini exhibited < 5% variance, while Copilot showed ≈ 8 − 12%; in all cases, hostility-aware weighting reduced the most aggressive expert’s influence by ≈ 27 − 29%, yielding robust group rankings. These mechanisms improve consensus quality by reducing bias from aggressive discourse, and they are expected to foster higher group satisfaction through perceived fairness in deliberation. Potential improvements include benchmarking against gold standards, extending to multilingual and multimodal contexts, and enhancing transparency for end-users. This work has been supported by the grant PID2022–139297OB-I00 funded by MICIU/ AEI/10.13039/501,100,011,033 and by ERDF/EU. Moreover, it is part of the project C-ING-165-UGR23, co-funded by the Regional Ministry of University, Research and Innovation and by the European Union under the Andalusia ERDF Program 2021–2027. Funding for open access publishing: Universidad de Granada/CBUA.