Learning Interpretable Logic for Agent-Based Models from Domain Independent Primitives

Genetic programming (GP) is a powerful method applicable to Inverse Generative Social Science (IGSS) for learning non-trivial agent logic in agent-based models (ABMs). While previous attempts at using evolutionary algorithms for learning ABM structures have focused on recombining domain-specific primitives, this paper extends prior work by developing techniques to evolve interpretable agent logic from scratch using a highly flexible domain-specific language (DSL) comprised of domain-independent primitives, such as basic mathematical operators. We demonstrate the flexibility of our method by learning symbolic models in two distinct domains: flocking and opinion dynamics, targeting data generated by reference models. Our results show that the evolved solutions closely resemble the reference models in behavior, generalize exceptionally well, and exhibit robustness to noise. Additionally, we provide an in-depth analysis of the generated code and intermediate behaviors, revealing the training process's progression. We explore techniques for further enhancing the interpretability of the resulting code and include a population-level analysis of the diversity for both models. This research demonstrates the potential of GP in IGSS for learning interpretable agent logic in ABMs across various domains.