On the usefulness of interactive computer game logs for agent modelling

Abstract

Interactive computer game logs show the potential for use as replacement for time-consuming supervisory learning processes for embodied, situated agents. However, due to the inherent nature of the data in the logs themselves, for the time being this promise cannot be fulfilled. An unsuccessful attempt to use largely rule-based data mining processes to learn behaviours from game logs led to finding the inherently top-down nature of such processes was fundamentally at odds with the unsupervised bottom-up learning requirement of the problem. Innate issues with game logs toward the goal of unsupervised agent learning are discussed. Possible approaches to the problem subsuming successful applications of various methods in narrower fields are presented for both symbolic and sub-symbolic advocates. © 2008 Springer Berlin Heidelberg.