Towards manifold learning for gamebot behavior modeling

摘要

Traditionally Computer Game Agent behaviors are generated by top-down approaches like finite state machines or scripts. So far, however, this had only mediocre success in creating life-like impressions. The bottom-up approach of imitation learning for agents has become very popular in recent robotics research and, in earlier work, we already discussed how imitation learning may apply to the programming of life-like computer game characters. However, so far we ignored problems concerning high dimensional state spaces for the most part, although behavior execution and learning takes place in such spaces.In this paper, we investigate the usage of non-linear dimensionality reduction for gamedata. We therefore focus on the aspect of topological gameworld representations and their dimensionality reduced counterparts. Dimensionality reduction is achieved by learning manifolds using Locally Linear Embedding. A mapping between data and embedding space is realized by Radial Basis Function interpolators. Experiments focus on movement path calculation and comparison in 3D and 2D embedding space world representations. The results indicate certain problems inherent to this approach but nevertheless justify further investigations.