Analyzing sequential Graph Generation with Graph Convolutional Policy Networks

Abstract

One property of Graph Convolutional Neural Networks (GCN) is to be permutation equivariant with respect to the node ordering. When GCN are used in combination with Reinforcement Learning (RL) and the action space depends on the node labeling then the action space is not equivariant by default. In this work we empirically show on very small graphs that the Graph Convolutional Policy Network introduced in [1] cannot generalize to symmetries introduced by node permutations. We extend the method to deal with permutation symmetries by using representatives of an isomorphic class of valid constructed subgraphs for a desired graph structure.