Reinforcement Learning for Node Selection in Mixed Integer Programming

Mixed Integer Linear Programming (MILP) optimization is a critical NP-hard problem applied across various sectors. Although machine-learning-based strategies have used historical data and expert strategies to effectively expedite optimal solution identification, they often neglect the importance of early search tree exploration and the cost associated with path backtracking. In this paper, we propose an innovative reinforcement learning (RL) approach for the node selection in the Branch-and-Bound (B&B) process. Our solution includes a novel reward function, an epsilon-greedy strategy for balancing exploitation and exploration, and a unique RL technique adaptable to the dynamic nature of the state and action spaces. Our empirical results show that our approach achieves an average of 10% acceleration compared to traditional SCIP and significantly outperforms other AI-based methods, underscoring the immense potential of RL in MILP optimization.