Model-Based Planning in Discrete Action Spaces.

Planning actions using learned and differentiable forward models of the world is a general approach which has a number of desirable properties, including improved sample complexity over model-free RL methods, reuse of learned models across different tasks, and the ability to perform efficient gradient-based optimization in continuous action spaces. However, this approach does not apply straightforwardly when the action space is discrete, which may have limited its adoption. In this work, we introduce two discrete planning tasks inspired by existing question-answering datasets and show that it is in fact possible to effectively perform planning via backprop in discrete action spaces using two simple yet principled modifications. Our experiments show that this approach can significantly outperform model-free RL based methods and supervised imitation learners.