Robust experience replay sampling for multi-agent reinforcement learning

Abstract

• Propose new algorithms for acquiring suitable experiences from buffer through filtering. • Strengthen exploration strategy by reducing repetitive decisions at a given state. • Improve performance which is higher than or comparable to the baseline algorithms. • Achieve early convergence and improved policy searching compared to the baselines. Learning from the relevant experiences leads to fast convergence if the experiences provide useful information. We present the new and simple yet efficient technique to find suitable samples of experiences to train the agents in a given state of an environment. We intended to increase the number of states visited and unique sequences that efficiently reduce the number of states the agents have to explore or exploit. Our technique implicitly introduces additional strength to the exploration-exploitation trade-off. It filters the samples of experiences that can benefit more than half the number of agents and then utilizes the experiences to extract useful information for decision making. First, we compute the similarities between the observed state and previous states in the experiences to achieve this filtering. Then, we filter the samples using the hyper-parameter, z , to decide which experiences will be suitable. We found out that agents learn quickly and efficiently since sampled experiences provide useful information that speeds up convergence. In every episode, most agents learn or contribute to improve the total expected future return. We further study our approaches’ generalization ability and present different settings to show significant improvements in diverse experiment environments.