Learning environment simulators from sparse signals

Abstract

To allow planning in novel environments that have not been mapped out by hand, we need ways of learning environment models. While conventional work has focused on video prediction as a means for environment learning, this work instead seeks to learn from much sparser signals, like the agent's reward. In Chapter 1, we establish a taxonomy of environments and the attributes that make them easier or harder to model through learning. In Chapter 2, we review prior work in the field of environment learning. In Chapter 3, we propose a model-learning architecture based purely on reward prediction, and analyze its performance on illustrative problems. Finally, in Chapter 4, we propose and evaluate a model-learning architecture that uses both reward and sparse "features" extracted from the environment.