A Learning based Branch and Bound for Maximum Common Subgraph Problems

no code implementations • 15 May 2019 • Yan-li Liu, Chu-min Li, Hua Jiang, Kun He

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.

reinforcement-learning Reinforcement Learning (RL)