A Hybrid Reinforcement and Self-Supervised Learning Aided Benders Decomposition Algorithm

TL;DR

This paper introduces a hybrid reinforcement and self-supervised learning framework to accelerate generalized Benders decomposition, achieving significant solution time reductions in mixed integer nonlinear programming.

Contribution

It combines reinforcement learning and neural network predictions to improve Benders decomposition efficiency and accuracy.

Findings

Achieves 57.5% reduction in solution time compared to classical GBD.

Consistently recovers optimal solutions across all test instances.

Uses a graph-based RL agent and KKT-informed neural network for solution prediction.

Abstract

We propose a hybrid reinforcement and self-supervised learning framework for accelerating generalized Benders decomposition (GBD). In this framework, a graph based reinforcement learning agent operates on a bipartite representation of the master problem and, together with a verification mechanism, determines the integer variable assignments that solve the master problem. These assignments are then used as inputs to a KKT informed neural network, trained via self supervision to predict primal dual solutions that approximately satisfy the Karush Kuhn Tucker conditions of the subproblem. The predicted solutions are used to construct Benders cuts directly. The framework is evaluated on a mixed integer nonlinear programming case study, where it achieves a 57.5% reduction in solution time relative to classical GBD while consistently recovering optimal solutions across all test instances.