Neural Algorithms for Graph Navigation

Abstract

The application of deep reinforcement learning (RL) to graph learning and meta-learning admits challenges from both topics. We consider the task of one-shot, partially observed graph navigation, acknowledging and addressing the difficulties of partially observed graph environments. In this work, we present a framework for graph meta-learning, and we propose an agent equipped with external memory and local action priors adapted to the underlying graphs. We demonstrate the efficacy of our framework through partially-observed navigation on synthetic graphs, as well as application to partially-observed navigation on 3D meshes, showing substantially improvement in one-shot performance over baseline agents.