Scalable Agent-Based Simulations of Normative Human Behavior

Abstract

The COVID-19 pandemic revealed critical limitations in the models that governments and researchers used to guide policy under uncertainty. Conventional compartmental models, such as S(E)I(A)R, treat populations as homogeneous and largely ignore behavioral and spatial heterogeneity, social influence, and non-uniform compliance with public measures. Yet, the effectiveness of behavioral interventions, like mask mandates, social distancing or lockdowns, depends as much on individual and social responses as on the properties of the disease. This thesis addresses the methodological challenge of capturing these aspects in computational simulations through a scalable agent-based framework that can deliver behavioral realism and normative (social) reasoning at scale. The thesis begins by reviewing the state of the art in agent-based modeling and simulation (ABMS), emphasizing the trade-off between behavioral realism and computational scalability. The literature shows a split between models that employ BDI models that support rich behavior at limited scale, and models that scale well but with limited behavioral complexity. To bridge these gaps, four major technical contributations are made: * Sim-2APL - a library that extends the 2APL BDI agent programming language with time-step synchronization and deterministic scheduling, enabling integration with external simulation environments. Sim-2APL supports modular (normative) agent reasoning allowing agents to weigh institutional mandates and social pressures in their behavioral decisions. It provides reusable design patterns and scales across distributed systems. * Sim-2APL+PanSim combines Sim-2APL with the distributed epidemic simulation environment {\em PanSim}. This combined framework represents the first platform that can scale simulations to millions of agents with rich behavior and autonomous (normative) reasoning. Using real data from Virginia (USA), the combined system reproduces the first six months of the COVID-19 pandemic, demonstrating both scalability and calibration against empirical mobility and infection data. Through counterfactual simulations, we explore the consequences of different compliance patterns and intervention strategies, illustrating how small changes in the intervention strategy, as well as its adoption by the population, can dramatically alter epidemic outcomes. * A Bayesian Optimization Framework for Policy Evaluation leverages this simulation framework to search efficiently for behavioral intervention strategies that balance epidemiological outcomes and societal costs. The approach demonstrates how simulation-based optimization can support policymaking by identifying robust interventions under the uncertaintanty of the adoption of that policy in the general population. * GenSynthPop - a method for constructing synthetic populations from aggregated statistical data, overcomes limitations of existing sample-based or stochastic approaches. It combines spatially detailed marginal distributions with higher-level contingency tables to deterministically generate realistic, spatially explicit populations of individuals and households, preserving key attribute correlations. This ensures that simulation agents reflect realistic socio-demographic interdependencies even where microdata are unavailable. Together, these components constitute a unified, scalable framework for data-driven behavioral simulation. The framework integrates realistic decision-making, norm compliance, social influence, and demographic heterogeneity within computationally tractable models. While demonstrated in epidemiology, the methodology extends to a wide range of policy domains, such as urban mobility, sustainability transitions, health promotion, and disaster response, where collective outcomes depend on adaptive human behavior, that may be nudged in a certain direction through policy.