Code: Learning Incentive Structures for Cooperative Resilience in Multi-Agent Systems under Social Dilemmas

This repository accompanies the paper "Learning Reward Functions for Cooperative Resilience in Multi-Agent Systems", which investigates how reward function design impacts cooperative resilience in Multi-Agent Reinforcement Learning (MARL). In dynamic and failure-prone environments, agents must not only optimize individual objectives but also ensure the collective system remains functional under disruptions. We define cooperative resilience as the ability of agents to anticipate, resist, recover, and adapt in the presence of external shocks. This repository provides tools and experiments to study and improve this emergent property through IRL-guided reward learning. We introduce a novel reward learning framework that learns reward functions from ranked trajectories—evaluated via a cooperative resilience score. Agents are then trained in social dilemma environments using: (i) Traditional individual reward functions (ii) Inferred rewards aligned with cooperative resilience (iii) Hybrid rewards combining both The reward inference is performed using two preference-based IRL algorithms across three types of parameterizations: Handcrafted features Linear reward models Neural networks Our results show that resilience-guided rewards lead to improved robustness and coordination, helping agents avoid catastrophic outcomes (e.g., resource depletion), without sacrificing individual performance. We further extend the experiments to larger 16×16 environments with 4 agents and multiple resource clusters, and evaluate agents under three disruption protocols (resource removal, regeneration slowdown, and agent perturbation).