Genetic Programming (GP) evolves computer programs to solve problems while showing promising avenues for discovering complex behavioural rules in agent-based models (ABMs). Since existing efforts to learn ABM structures in the field of Inverse Generative Social Science (IGSS) have concentrated on combining domain-specific primitives for deterministic rule generation, this paper evolves interpretable agent logic from scratch comprised of stochastic primitives for decision tree nodes. We show the adaptability of our approach by applying it to discover models representing human behaviour targeting data generated from existing conceptual models. Our findings demonstrate that IGSS successfully identified the reference model’s stochastic behaviour, with six of the top ten evolved rules matching the age-based random exit selection strategy. However, permutation importance analysis revealed that distance-based exit selection and gender similarity metrics demonstrated higher significance than age-based parameters, despite the latter being fundamental to the reference model. This apparent discordance can be attributed to the stochastic initialisation of agent attributes within the ABM simulation environment. While consistent convergence was observed by the fifth generation, the presence of substantial fitness fluctuations highlighted inherent challenges in learning from noisy data. This emerges as a critical challenge in stochastic ABM model discovery, necessitating the development of novel methodologies to guide the evolution of tree-based rules in genetic algorithms that are robust against noise-induced variations.