Predictive Projections

摘要

This paper addresses the problem of learning control policies in very high dimensional state spaces. We propose a linear dimensionality reduction algorithm that discovers predictive projections: projections in which accurate predictions of future states can be made using simple nearest neighbor style learning. The goal of this work is to extend the reach of existing reinforcement learning algorithms to domains where they would otherwise be inapplicable without extensive engineering of features. The approach is demonstrated on a synthetic pendulum balancing domain, as well as on a robot domain requiring visually guided control.