Improving Branch-and-Bound Using Decision Diagrams and Reinforcement Learning

Abstract

Combinatorial optimization has found applications in numerous fields, from transportation to scheduling and planning. The goal is to find an optimal solution among a finite set of possibilities. Most exact approaches use relaxations to derive bounds on the objective function, which are embedded within a branch-and-bound algorithm. Decision diagrams provide a new approach for obtaining bounds that, in some cases, can be significantly better than those obtained with a standard linear programming relaxation. However, it is known that the quality of the bounds achieved through this bounding method depends on the ordering of variables considered for building the diagram. Recently, a deep reinforcement learning approach was proposed to compute a high-quality variable ordering. The bounds obtained exhibited improvements, but the mechanism proposed was not embedded in a branch-and-bound solver. This paper proposes to integrate learned optimization bounds inside a branch-and-bound solver, through the combination of reinforcement learning and decision diagrams. The results obtained show that the bounds can reduce the tree search size by a factor of at least three on the maximum independent set problem.KeywordsDecision diagramsBranch-and-boundReinforcement learning.