Multiagent Safe Reinforcement Learning by Decentralized Event-Triggered Min-Max Optimization
Shunsuke Otani, Naoki Hayashi, Masahiro Inuiguchi · 2025
This paper addresses decentralized event-triggered reinforcement learning with safety constraints. Each agent has an individual reward function and safety constraints that depend on the joint actions of agents. The objective is to maximize the team’s long-term return while satisfying the safety constraints. We formulate the reinforcement learning problem as a nonconvex-concave min-max optimization problem and propose a decentralized policy gradient algorithm. Each agent has estimations for the optimal primal and dual solutions of the min-max optimization problem. Different from the existing decentralized reinforcement learning, the agents share these estimates only when the error exceeds a predefined threshold. We show that the estimates of agents converge to a neighborhood of a locally optimal solution while effectively reducing the communication overhead.