Deep Learning × ABM
this research investigates how deep learning — specifically recurrent and graph neural networks — can augment agent-based models (ABMs) to improve their ability to simulate and predict complex adaptive systems. accepted to the European Social Simulation and Economics conference in Ghent, Belgium.
the limitation of ABMs
agent-based models are powerful for simulating emergent phenomena: economic dynamics, epidemics, opinion formation, market crashes. but calibrating them to real-world data is notoriously difficult. you can specify agent rules, but ensuring those rules reproduce empirically observed behavior requires either domain expertise (introducing bias) or computationally expensive optimization. most ABMs are calibrated manually, which limits their predictive accuracy.
what deep learning adds
neural networks, particularly recurrent and graph architectures, are well-suited for learning complex behaviors from data trajectories. the research explores three contributions: (1) learning agent behavior rules directly from observed data rather than specifying them by hand, (2) accelerating ABM simulations through neural surrogate models that approximate the full simulation at a fraction of the computational cost, and (3) improving calibration through differentiable simulation frameworks.
ESEE 2026
the paper was accepted to the European Social Simulation and Economics conference in Ghent, Belgium — one of the leading interdisciplinary venues for computational social science and agent-based modeling. the contribution proposes a framework for integrating learned agent behaviors with interpretable simulation, with applications to economic modeling, epidemiology, and policy impact analysis.