Scalable and cohesive swarm control based on reinforcement learning

Abstract

Unmanned vehicles have seen a significant increase in a wide variety of fields such as for logistics, agriculture and other commercial applications. Controlling swarms of unmanned vehicles is a challenging task that requires complex autonomous control systems. Reinforcement learning has been proposed as a solution to this challenge. We propose an approach based on agent masking to enable a simple Deep Q-Network algorithm to scale on large swarms while training on relatively smaller swarms. We train our approach using multiple swarm sizes and learning rates and compare our results using metrics such as the number of collisions. We also compare the ability of our approach to scale on swarms ranging from five to 25 agents using metrics and visual analysis. Our proposed solution was able to guide a swarm of up to 100 agents to a target while keeping a good swarm cohesion and avoiding collision.