Deep Auto-Deferring Policy for Combinatorial Optimization

Designing efficient algorithms for combinatorial optimization appears ubiquitously in various scientific fields. Recently, deep reinforcement learning (DRL) frameworks have gained considerable attention as a new approach: they can automatically learn the design of a good solver without using any sophisticated knowledge or hand-crafted heuristic specialized for the target problem. However, the number of stages (until reaching the final solution) required by existing DRL solvers is proportional to the size of the input graph, which hurts their scalability to large-scale instances. In this paper, we seek to resolve this issue by proposing a novel design of DRL's policy, coined auto-deferring policy (ADP), automatically stretching or shrinking its decision process. Specifically, it decides whether to finalize the value of each vertex at the current stage or defer to determine it at later stages. We apply the proposed ADP framework to the maximum independent set (MIS) problem, a prototype of NP-complete problems, under various scenarios. Our experimental results demonstrate significant improvement of ADP over the current state-of-the-art DRL scheme in terms of computational efficiency and approximation quality. The reported performance of our generic DRL scheme is also comparable with that of the state-of-the-art solvers specialized for MIS, e.g., ADP outperforms them for some graphs with millions of vertices.