Collaborative Training

Ariana Mirella Villegas Suarez

Track: Short Paper

Abstract: We present a framework for a parameter-sharing mechanism based on multi-agent reinforcement learning. Our approach allows agents to balance exploration and exploitation, sharing parameters only when a significant performance gap is detected. Experiments conducted across six environments show that our framework achieves up to 40% faster convergence and improves cumulative rewards by 15% in complex tasks. In addition, we observe a 25% reduction in performance variance among agents, showing the robustness and efficiency of our collaborative strategy.

Submission Number: 62