Resilient Multi-agent Reinforcement Learning Using Medoid and Soft-medoid Based Aggregation

A network of reinforcement learning (RL) agents that cooperate with each other by sharing information can improve learning performance of control and coordination tasks when compared to non-cooperative agents. However, networked Multi-agent Reinforcement Learning (MARL) is vulnerable to adversarial agents that can compromise some agents and send malicious information to the network. In this paper, we consider the problem of resilient MARL in the presence of adversarial agents that aim to compromise the learning algorithm. First, the paper presents an attack model which aims to degrade the performance of a target agent by modifying the parameters shared by an attacked agent. In order to improve the resilience, the paper presents aggregation methods using medoid and soft-medoid. Our analysis shows that the medoid-based MARL algorithms converge to an optimal solution given standard assumptions, and improve the overall learning performance and robustness. Simulation results show the effectiveness of the aggregation methods compared with average and median-based aggregation.