Sociodynamics of Reinforcement Learning

Sociodynamics of Reinforcement Learning Yann Bouteiller, Karthik Soma, Giovanni Beltrame Accepted to Transactions on Machine Learning Research (TMLR); presented at RLC 2026 We study the collective behavior of large populations of co-learning reinforcement learning agents — up to 200,000 naive and opponent shaping (LOLA) learners — paired through evolutionary-style simulations. We show that these population-level dynamics encode rich information that mean-field or self-play approximations often fail to capture, a small step toward a "psychohistory" (from Asimov's Foundation) of large agent societies. 📄 Paper: https://arxiv.org/abs/2410.17466 💻 Code: https://github.com/MISTLab/RL-societies #MultiAgentRL #LOLA #Alife