Online Learning From Observation For Interactive Computer Games

Abstract

The research presented in this paper describes an architecture, which enables an agent to predict an observed entity’s actions (most likely a human’s) online. Case-based approaches have been utilised by a number of researchers for online action prediction in interactive applications. Our architecture builds on these works and provides a number of novel contributions. Specifically our architecture offers a more comprehensive state representation, behaviour prediction and a more robust case maintenance approach. The proposed architecture is fully described in terms of interactive simulations (specifically first person shooter (FPS) computer games); however it would be applicable to other interactive applications, such as intelligent tutoring and surveillance systems. We conclude the paper by evaluating our proposed architecture and discussing how the system will be implemented.