Aprendizaje Automático para Modelado Basado en Agentes. Diseño Inductivo del Comportamiento de los Agentes

• Autores: Víctor Alejandro Vargas Pérez
• Directores de la Tesis: Óscar Cordón García (codir. tes.) , Manuel Chica Serrano (codir. tes.)
• Lectura: En la Universidad de Granada ( España ) en 2025
• Idioma: español
• ISBN: 9791370150488
• Número de páginas: 247
• Tribunal Calificador de la Tesis: Luciano Sánchez Ramos (presid.) , Raúl Pérez Rodríguez (secret.) , Alicia Troncoso Lara (voc.) , Rocío Celeste Romero Zaliz (voc.) , Raymond Chiong (voc.)
• Programa de doctorado: Programa de Doctorado en Tecnologías de la Información y la Comunicación por la Universidad de Granada
• Menciones:
  • Internacional
• Calificación: Sobresaliente Cum Laude
• Enlaces
  • Tesis en acceso abierto en: DIGIBUG
• Resumen

  • Complex social systems, which are characterized by nonlinear interactions among myriad actors, can quickly amplify local actions into large-scale, often unpredictable dynamics. Understanding and ultimately steering such behavior requires computational frameworks capable of bridging micro-level interactions with the macro-level patterns they give rise to. Two widely used approaches are Agent-Based Models (ABMs), which reconstruct the system from the bottom-up by explicitly simulating autonomous agents and their interactions, and Social Network Analysis (SNA), a descriptive framework that examines the structural properties of the underlying interaction network. Recent advances in Machine Learning (ML) and the increase of data availability invite a more inductive perspective: instead of hand-coding agent rules or tuning parameters case by case, we can let data guide rule design, parameters calibration, and patterns extraction. This thesis installs ML as a third methodological pillar alongside ABMs and SNA, arguing that the triad enables richer explanation and decision support. Four contributions build that vision across marketing, public opinion dynamics, and Peer-to-Peer (P2P) accommodation markets. Our first contribution starts from a marketing ABM calibrated with real-world data, in which it integrates a Double Deep Q-Network (DDQN) agent representing a brand with the aim of learning a dynamic media planning policy which maximizes brand awareness under budget constraints. Tested on seven scenario variants, the learned policy beat historic spending plans in every case, raising awareness by 1–19 percentage points and, in the most realistic scenario, from 44.7% to 64.5%. It also outperformed an econometric baseline, while its spending patterns aligned intuitively with channel effectiveness: concentrated where touchpoints impact differed widely, diversified when impacts were similar. These results demonstrate how an ABM sandbox can be upgraded into a proactive policy recommendation agent. ABMs for opinion dynamics often assume uniform confidence thresholds because individuallevel values are unobserved, and automatic calibration for a full heterogeneous population is costly and non-transferable. The second study frames threshold estimation as supervised learning on graphs. Synthetic training data pairs diverse social topologies (random, smallworld, and scale-free social networks) with initial-and-final opinion states generated by the Hegselmann–Krause model. Graph neural networks then learn to regress a scalar threshold per node. Tested on five real topologies of up to twenty times the training size, the best model (GraphSAGE) achieves a R"2 > 0.7 across all cases and, crucially, transfers to unseen, larger social networks without retraining, enabling one-shot estimation of fully heterogeneous thresholds without further ABM simulations. Real-world opinions rarely converges smoothly but rather exhibit oscillating dynamics. Study 3 injects that volatility into three canonical opinion dynamics models (Friedkin– Johnsen, Deffuant–Weisbuch, and ATBCR) by allowing parameters to change monthly. A scale-free network with 3 961 nodes represents the January 2023 Spanish CIS survey, and initial concern for immigration is derived from respondents’ ranked problem lists. Four evolutionary algorithms (DE, SHADE, L-SHADE, and PSO) explore 30 000 parameter solutions per model, minimizing the output’s deviation against 15 monthly barometers (January 2023 – April 2024). Only the hybrid ATBCR model reproduces the observed peaks and troughs, delivering parameter values with an interpretable narrative: concern peaks coincide with agents prone to emotional polarization, while troughs with rational consensus mechanisms. Thus, the resulting parameter trajectories offer a quantifiable narrative of the public conversation mechanism shifts which domain experts can interpret and link to real-world events. The fourth and last contribution develops a visual decision support system for heterogeneous P2P accommodation markets, where investment success hinges on identifying profitable mixes of property attributes. We construct a Pathfinder relational network based on a dataset of more than 9 000 Airbnb listings from Spain’s Canary Islands (November 2020 – October 2021). The network’s nodes represent listings and its edges encode HEOM-based mixed-type similarity. Redundant links are pruned while preserving the network’s global structure. A force-directed layout (OpenOrd followed by Yifan Hu) and Louvain community detection reveal nine well-separated clusters. Overlaying exogenous economic performances (absent from the similarity metric) uncovers a pronounced core-periphery pattern: highly profitable, large-capacity properties populate an outer ring section of the generated layout, whereas average listings crowd the center. Furthermore, superhost status and flexible cancellation further increase returns, while professional multi-listing hosts gravitate to mid- or low-profit clusters, echoing evidence that host scale can dilute service quality. Overall, this network-based visual analysis offers actionable insights and concrete guidance for investors. Collectively, the studies show that ML can turn calibrated ABMs into real-time policy optimizers and recover heterogeneous agent attributes at scale, that embedding and calibrating temporal parameters into opinion dynamics models can explain the conversation mechanisms shifts behind oscillating public opinions, and that effective Pathfinder relational networks visualizations together with community detection and analysis can provide valuable insights into heterogeneous P2P markets. Future work will extend reinforcement learning application to multi-brand competition, apply graph neural networks inference to epidemiology and innovation diffusion, and compare Pathfinder-based segmentation with traditional clustering across different domains. In short, by weaving together ABM, SNA and ML, the thesis contributes to the development of data-driven toolkits for diagnosing and intervening in complex social systems.