Neighbor-Based Decentralized Training Strategies for Multi-Agent Reinforcement Learning
Date:
I presented our paper “Neighbor-Based Decentralized Training Strategies for Multi-Agent Reinforcement Learning” at SAC 2025.
The talk compared decentralized training strategies in which reinforcement-learning agents cooperate by exchanging experience or model information only with nearby peers. The study investigated experience sharing, neighbour-based averaging, and consensus as alternatives to fully centralized training.
The presentation emphasized how local interaction structures can support scalable cooperative learning while preserving the autonomy of individual agents.