Learn Branching Policies with an Attention-based Model

Abstract

The branch-and-bound method is an exact tree search method in mixed integer linear programming. Recently, learning branching policies become a popular topic, but most existing machine learning methods only consider problems with similar structures, and pay little attention to heterogeneous problems. The parameterized search tree method demonstrates promising performance on heterogeneous problems, but there is still room for improvement when it generalizes to other problems. In this paper, we propose an attention-based branching framework, which can exploit information among candidate variables. In this way, we can take full advantage of the properties exploited from the parameterized branch-and-bound search tree, resulting in a better branching policy. Our proposed model is evaluated on multiple mixed integer linear programming instances and attains remarkable performance within a given time limit.