Neural Networks in RTS AI

Student: Timothy Adam Smith

Supervisor: Mrs. Peta Wyeth

Category: Software Engineering Thesis Project

When playing Real Time Strategy Games (such as StarCraft and Age of Empires) human oppoents are more fun to play against than the computer players. This thesis investigates possibilites for making computers players play more like humans. Specifically, it looks at using a neural network with stochasticly interpreted outputs as a key portion of the AI.

The aim of using a Neural Network like this is to: avoid the vulnerabilities due to predictability that occur in some other game's AI's; allow the automatic discovery of tactics to choose from; facilitate the development of multiple distinct computer players and proved a challanging opponent without cheating (that is, without providing the computer opponent with resources or information that the human does not receive).

To this end a simple real-time strategy game, SURTS (shown in Figure 1), was written. Strategies were then evolved to play the game, with their fitnesses calcualted by playing against each other.