Q-learning in collaborative multiagent systems

Abstract

Q-learning is one of the most widely used reinforcement learning techniques. It is very effective for learning an optimal policy in any finite Markov decision process (MDP). Collaborative multiagent systems, though, are a challenge for self-interested agent implementation, as higher utility can be achieved via collaboration. To evaluate the Q-learning efficiency in collaborative multiagent systems, we will use a simplified version of the Malmo Collaborative AI Challenge (MCAC). It was designed by Microsoft and consists of a game where 2 players can collaborate to catch the pig (high reward) or leave the game (low reward). Each action costs 1, so knowing when to leave and when to chase the pig is key for achieving high scores. Two main problems are faced in the challenge: uncertainty of the other agent behaviour and a limited learning time. We propose solutions to both problems using a simplified MCAC environment, a stateaction abstraction and an agent type modelling. We have implemented an agent that is able to identify the other player behaviour (whether it is collaborating or not) and can learn an optimal policy against each type of player. Results show that Q-learning is an efficient and effective technique to solve collaborative multiagent systems.