Behaviour selection using neural networks

Artificial Neural networks (ANNs) have been successfully used in a handful of commercial game success to date, such as Lionhead Studios' Black and White [Evans]. The technology is still shrouded in mystery for many developers and the perceived complexity of implementation halts further take up of the technology. The experiment described in this paper gives a brief background to neural networks and applies the backpropagation neural network architecture to a simple agent behavior selection controller, which could equally have been implemented with a finite state machine or similar technology. The procedure and results of the experiment are discussed in relation to the general applicability of ANNs to game environments compared to the more traditional methods such as finite state machines.