Arelative position descriptor for multi-agent reinforcement learning

Since there exists cooperation or competition between the agents in multi-agent systems, relative position based information becomes very important for agents to learn collaborative behaviors. In this work, we propose a relative position descriptor to present the neighborhood relations of each agent. The relative position descriptor divides the neighborhood space of an agent into several bins according to the bearing angle, and the bins are utilized to store task-related information. The length of the relative position descriptor is fixed, so it is well suited for scenarios where the number of agents is large or changeable. We evaluate the proposed relative position descriptor on several multi-agent tasks. The experimental results illustrate that the relative position descriptor can improve the performance of multi-agent reinforcement learning.