Multi-Agent Reinforcement Learning in General Filling Problem

Abstract

In this study, a top-down Commander AI-Agent (CAIA) model and a bottom-up Multi-Agent Communication (MAC) model are applied to the "Generalized Filling Problem". This problem is researched as a game that requires multiple agents to fill an entire grid map and can be used to explore algorithms for coordinating multiple robots, such as cleaning robots, mine clearance, and drone control. Each agent has a neural network, the inputs of which are the coordinates and map information of each agent, and the outputs of which agent’s actions. The NN in the CAIA model is held by the Commander and a NN of agent in the MAC model is held by each agent. When the entire map is filled, rewards are given to all agents and the learning process is made progress. The goal of study is to compare the two models and determine which algorithm is more efficient, the CAIA model or the MAC model, depending on the shape of the map. A useful application of this research is as a way for multiple drones to efficiently explore a space.