Optimizing the calibration of agent-based models throughreinforcement learning

The calibration of agent-based models (ABMs) in economics and finance typically involves searching for “satisfactory” parameter values over high-dimensional parameter spaces, a process that is both computationally intensive and highly sensitive to the choice of search algorithm. To address these challenges, we introduce a reinforcement learning (RL) method that adaptively selects and combines multiple search heuristics throughout the calibration process. We evaluate the proposed approach using the macroeconomic ABM developed by Assenza et al.(2015)(ADG) and demonstrate that it consistently outperforms both individual algorithms and conventional hybrid benchmarks. The method is further assessed on models from finance and epidemiology, highlighting its robustness and applicability across domains. Finally, we calibrate and validate the ADG model against empirical moments and examine its dynamic responses to structural shocks, comparing the resulting outcomes across alternative calibration strategies.