Study of Multiple Multiagent Reinforcement Learning Algorithms in Grid Games
Pascal De Beck-Courcelle · 2013
This thesis studies multiagent reinforcement learning algorithms and their implementation; In particular the Minimax-Q algorithm, the Nash-Q algorithm and the WOLF-PHC algorithm.We evaluate their ability to reach a Nash equilibrium and their performance during learning in general-sum game environments.We also test their performance when playing against each other.We show the problems with implementing the Nash-Q algorithm and the inconvenience of using it in future research.We fully review the Lemke-Howson algorithm used in the Nash-Q algorithm to find the Nash equilibrium in bimatrix games.We find that the WOLF-PHC is a more adaptable algorithm and it performs better than the others in a general-sum game environment.