Evaluating Robustness of Cooperative MARL

Abstract

In recent years, a proliferation of methods were developed for multi-agent reinforcement learning (MARL). In this paper, we focus on evaluating the robustness of MARL agents in continuous control tasks. In particular, we propose the first model-based approach to perform adversarial attacks for cooperative MARL. We design effective attacks to degrade the MARL agent's performance by adversarially perturbing the states of agent(s) and solving an optimization problem. In addition, we also developed several strategies to select the most vulnerable agents that help to further decrease the team reward of MARL. Extensive numerical experiments on multi-agent Mujoco tasks verify the effectiveness of our proposed approach.

Tasks

Reproductions