Field-informed Reinforcement Learning of Collective Tasks with Graph Neural Networks

Coordinating a multi-agent system of intelligent situated agents is a traditional research problem, impacted by the challenges posed by the very notion of distributed intelligence. These problems arise from agents acquiring information locally, sharing their knowledge, and acting accordingly in their environment to achieve a common, global goal. These issues are even more evident in large-scale collective adaptive systems, where agent interactions are necessarily proximity-based, thus making the emergence of controlled global collective behaviour harder. In this context, two main approaches have been proposed for creating distributed controllers out of macro-level task/goal descriptions: manual design, in which programmers build the controllers directly, and automatic design, which involves synthesizing programs using machine learning methods. In this paper, we consider a new hybrid approach called Field-Informed reinforcement learning (FIRL). We utilise manually designed computational fields (globally distributed data structures) to manage global agent coordination. Then, using Deep Q-learning in combination with Graph Neural Networks we enable the agents to learn the necessary local behaviour automatically to solve collective tasks, relying on those fields through local perception. We demonstrate the effectiveness of this new approach in simulated use cases where tracking and covering tasks for swarm robotics are successfully solved.