Machine learning for real-time strategy computer games

A commercial Real-Time Strategy (RTS) game requires an artificial intelligence component, simply called an "AI", capable of providing a human player with a challenging opponent. These AIs must simulate the play style of a competent human player. The AIs of current RTS games can play with multiple levels of skill. However, their simulation of a competent human player is incomplete because they cannot change their tactics to adapt to a human player.;In this thesis, neural networks provide learning mechanisms for RTS AIs. Two AI prototypes that each learn a binary decision function are described. Both of these prototypes use a Stochastic Back-Propagation (SBP) algorithm to train a neural network. However, the data set for training the neural network is created in two different ways. The Moving Window technique records <input state, output state> tuples for each AI-controlled game agent whenever it makes a decision during games. When learning is performed, these tuples are evaluated for their contribution to the overall performance of the AI-controlled agents. The Simulated Memory technique extracts a set of <input state, output state> tuples from the neural network, updates this set based on the effectiveness of recent commands, and retrains the neural network with the updated set.;The two prototypes were tested against several scripted AI opponents. Various input parameters were tried to tune the performance of the prototypes. Both prototypes successfully adapted their strategies until they were victorious over simulated opponents with simple deterministic strategies. The Moving Window prototype performed better than the Simulated Memory technique with respect to accuracy and speed in the experiments for the parameter settings used.