Reviews: Graph Convolutional Policy Network for Goal-Directed Molecular Graph Generation

The paper proposes combining graph neural nets, RL, and adversarial learning to train a molecule generation procedure optimized according to specific task requirements. RL is used to optimize non-differentiable objectives. Adversarial learning is used to encourage the policy to generate molecules that are similar to known ones. Graph neural nets are used to capture the graph structure of the molecules. Results are shown for three different tasks -- optimizing a specific property of the generated molecule, generating molecules that satisfy constraints on property scores, and generating molecules that satisfy constraints on their composition.