Improved Learning in Multi-Agent Pursuer-Evader UAV Scenarios via Mechanism Design and Deep Reinforcement Learning

Cooperation is a challenging task as balancing the desired outcome for a group of autonomous agents while they perceive/act based on their own point of view poses an interesting research question. Many approaches have been proposed for dealing with this issue using Deep Reinforcement Learning (DRL) solutions, where the agents learn how to weigh their decisions based on the group's interests. In such solutions, the design of the reward functions for multi-agent DRL problems is an area that is still open to be explored given the complexity of the analysis and the environment-agents relationships. In this work, we propose the application of the Vickrey-Clarke-Groves mechanism design principles to derive reward functions to be applied during the training of autonomous agents in a cooperative multiple pursuer-evader game scenario. Such economic theory is very promising given the capability of incentivizing and penalizing agent's behaviour given the impact of their actions on the whole group. Our results show that the proposed reward scheme causes the agents to learn faster and to be more effective in achieving solutions to the pursuer-evader problem when compared to DRL solutions using traditional reward functions.