Towards Learned Anticipation in Complex Stochastic Environments

Abstract

I describe a novel methodology by which a software agent can learn to predict future events in complex stochastic environments. It is particularly relevant to environments that are constructed specifically so as to be able to support highperformance software agents, such as video games. I present results gathered from a first prototype of our approach. The technique presented may have applications that range beyond improving agent performance, in particular to user modeling in the service of automated game testing.