HEC Montreal

Published purpose

The objective of this project is to develop general machine learning techniques for graph generation, with the end application of smart design including new material discovery, advanced circuit design, and novel drug invention, amongst many others. Research will focus on deep generative models and reinforcement learning for the generation of graphs with optimized properties. The representation power of graph will be leveraged to sufficiently encode the key compositional behaviours and their interplays of the target domain, and treat developing a novel design as a new graph structure generation process with various composition constraints. The decomposition in the former can be attained through deep generative models with disentangled latent variables, and the composition search space in the latter can be effectively explored by deep enforcement learning.