Multi-agent Polygon Formation using Reinforcement Learning

This work provides details a simulation experiment and analysis of Q-learning applied to multi-agent systems. Six agents interact within the environment to form hexagon, square and triangle, by reaching their specific goal states. In the proposed approach, the agents form a hexagon and the maximum dimension of this pattern is be reduced to form patterns with smaller dimensions. A decentralised approach of controlling the agents via Q-Learning was adopted which reduced complexity. The agents will be able to either move forward, backward and sideways based on the decision taken. Finally, the Q-Learning action-reward system was designed such that the agents could exploit the system which meant that they would earn high rewards for correct actions and negative rewards so the opposite.