Emergence of Game Strategy in Multiagent Systems

In this thesis we focused on subsymbolic approach to machine game play problem. We worked on two different methods of learning. Our first goal was to test the ability of common feed-forward neural networks and the mixture of expert topology. We have derived reinforcement learning algorithm for mixture of expert network topology. This topology is capable to split the problem into smaller parts, which are easier to be solved by an expert neural network. We have compared the quality of strategy emergence between mixture of expert networks and feed-forward networks. Our experiments demonstrate that mixture of experts is able to play a game at the same level as feed-forward networks with equal number of weights. The second approach derived in this work is reinforcement learning with usage of extended Kalman filer. Extended Kalman filter can be used for neural network training. Its advantage is very high learning rate in terms of training cycles. We have proposed usage of extended Kalman filter for reinforcement learning with TD(0) and Monte Carlo method. We have compared the quality of strategy emergence between extended Kalman filter and TD(λ) approach. Our results show that extended Kalman filter is able to create a game strategy after playing a considerably fewer number of games.