An Adaptive Evolutionary Approach for Discovering Stochastic Agent-Based Models

This research proposes an Adaptive Genetic Programming (AGP) approach within the Inverse Generative Social Science (IGSS) framework to effectively discover stochastic behavioural rules for Agent-Based Models (ABMs). Our method explicitly incorporates stochastic decision-making alongside deterministic primitives to realistically simulate complex human behaviours, exemplified by pedestrian exit choices in crowded environments. The AGP algorithm integrates dynamic population resizing, elite-based restarts, and adaptive termination criteria, enhancing computational efficiency and robustness against local optima. Through rigorous evaluation, the AGP successfully recovered the original pseudo-truth rule—agents selecting exits based on combined distance and crowding—used to generate synthetic datasets. Notably, rules considering both crowd density and distance outperformed simpler rules relying solely on proximity. Robustness analyses demonstrated that the evolved pseudo-truth rule consistently achieved better performance compared to other evolved alternatives, while most evolved rules performed significantly better than the null comparator baseline. Sensitivity analysis further validated the algorithm's effectiveness in balancing exploration and computational cost. These results demonstrate AGP's potential for uncovering interpretable and empirically grounded behavioural rules, with broad applicability to various stochastic social simulation scenarios.