Emergence of Cooperation in Heterogeneous Populations of Reinforcement-Learning Agents

Emergence of Cooperation in Heterogeneous Populations of Reinforcement-Learning Agents

Mathematics

Daphne Lavi

2025

Paolo Turrini and Chin-wing Leung

This project extends the framework of Leung and Turrini (AAMAS 2024) by introducing behavioural diversity into populations of reinforcement learning agents playing repeated Prisoner's Dilemma games with partner selection. Agents differ in their learning rate and exploration-exploitation balance, forming “curious” and “conservative” types. Simulations show that heterogeneity sustains long-term cooperation. These insights contribute to understanding how diversity influences collective learning dynamics in multi-agent systems.

multi-agent systems, reinforcement learning, cooperation, social dilemmas, Prisoner’s Dilemma, partner selection, heterogeneity, exploration-exploitation, Q-learning, Boltzmann exploration, AI agents