IMITATIVE LEARNING OF COMBAT BEHAVIOURS IN FIRST-PERSON COMPUTER GAMES

Modern, commercial computer games rely primarily on AI techniques that were developed several decades ago, and until recently there has been little impetus to change this. Despite the fact that the computer-controlled agents in such games often possess abilities far in advance of the limits imposed on human participants, competent players are capable of easily beating their artificial opponents suggesting that approaches based on the analysis and imitation of human play may produce superior agents, both in terms of performance and believability. The very fact that games provide the ability to quickly and easily generate vast quantities of raw, objective human behavioural data presents many fascinating opportunities to the AI community; opportunities which, with few exceptions, have not yet been suitably explored. Through our work in the field of imitation learning, we therefore investigate how best to utilise the low-level data accrued from recorded game sessions in the creation of intelligent, convincingly human game agents. In previous contributions, we have described models capable of imitating goal-oriented strategic navigation (Gorman & Humphrys 2005) and of reproducing characteristically human movement in first-person shooter games (Gorman, Thurau, Bauckhage & Humphrys 2006a); we have also outlined a comprehensive approach to the evaluation of agent believability (Gorman et al 2006b). Here, we present an approach to the imitation of combat behaviours in such environments. We first describe the extraction and processing of relevant feature vectors from the game session using our custom-built QASE API (Gorman, Fredriksson & Humphrys 2005). We then outline a neural-network based model designed to learn the aiming and context-sensitive weapon handling exhibited by the human players. Finally, we describe an experiment to demonstrate the efficacy of this approach; some observations and future directions for our work close this contribution.