Oracle-SAGE: Planning Ahead in Graph-Based Deep Reinforcement Learning

Deep reinforcement learning (RL) commonly suffers from high sample complexity and poor generalisation, especially with high-dimensional (image-based) input. Where available (such as some robotic control domains), low dimensional vector inputs outperform their image based counterparts, but it is challenging to represent complex dynamic environments in this manner. Relational reinforcement learning instead represents the world as a set of objects and the relations between them; offering a flexible yet expressive view which provides structural inductive biases to aid learning. Recently relational RL methods have been extended with modern function approximation using graph neural networks (GNNs). However, inherent limitations in the processing model for GNNs result in decreased returns when important information is dispersed widely throughout the graph. We outline a hybrid learning and planning model which uses reinforcement learning to propose and select subgoals for a planning model to achieve. This includes a novel action selection mechanism and loss function to allow training around the non-differentiable planner. We demonstrate our algorithms effectiveness on a range of domains, including MiniHack and a challenging extension of the classic taxi domain.