GCT-MARL Boosts Multi-Agent RL Transfer Learning.

Training agents for cooperative multi-agent reinforcement learning (MARL) from scratch for every new environment or task is often computationally expensive and time-consuming, posing a significant challenge for real-world deployment. This research introduces GCT-MARL, a novel transfer learning framework designed to enhance sample efficiency in MARL. GCT-MARL builds upon the multi-view graph contrastive backbone of MAIL, augmenting it with a per-view, adaptively weighted alignment loss. It also incorporates a two-phase training protocol specifically tailored for effective transfer across agent populations that vary in size and composition. Empirical evaluations demonstrate that GCT-MARL markedly accelerates convergence on target tasks compared to training from scratch. This improvement is observed in both homogeneous transfer scenarios (within the same agent faction but with varying numbers of agents) and heterogeneous scenarios (across different factions and mixed unit types). Furthermore, the framework naturally supports continual learning by allowing the two-phase transfer protocol to be chained sequentially across a series of related tasks. This work provides a unified approach to mitigate key limitations in current MARL transfer methods.