BAYESIAN MODELS FOR KEYHOLE PLAN RECOGNITION IN AN ADVENTURE GAME

We present an approach to keyhole plan recognition which uses a dynami c belief (Bayesian) network to represent features of the domain that a re needed to identify users' plans and goals. The application domain i s a Multi-User Dungeon adventure game with thousands of possible actio ns and locations. We propose several network structures which represen t the relations in the domain to varying extents, and compare their pr edictive power for predicting a user's current goal, next action and n ext location. The conditional probability distributions for each netwo rk are learned during a training phase, which dynamically builds these probabilities from observations of user behaviour. This approach allo ws the use of incomplete, sparse and noisy data during both training a nd testing. We then apply simple abstraction and learning techniques i n order to speed up the performance of the most promising dynamic beli ef networks without a significant change in the accuracy of goal predi ctions. Our experimental results in the application domain show a high degree of predictive accuracy. This indicates that dynamic belief net works in general show promise for predicting a variety of behaviours i n domains which have similar features to those of our domain, while re duced models, obtained by means of learning and abstraction, show prom ise for efficient goal prediction in such domains.