A Learning based Branch and Bound for Maximum Common Subgraph Problems

Abstract: Branch-and-bound (BnB) algorithms are widely used to solve combinatorial problems, and the performance crucially depends on its branching heuristic.In this work, we consider a typical problem of maximum common subgraph (MCS), and propose a branching heuristic inspired from reinforcement learning with a goal of reaching a tree leaf as early as possible to greatly reduce the search tree size.Extensive experiments show that our method is beneficial and outperforms current best BnB algorithm for the MCS.