Resumen de Neural benders decomposition for mixed-integer programming
Rahimeh Neamatian Monemi, Shahin Gelareh, Nelson Maculan, Hong Dai
In this study, we propose an imitation learning framework to enhance the Benders decomposition method. This work aims to learn how to select dual values when there is a choice to be made among alternatives. To attain this objective, we mimic successful experts via two policies. In the first one, we replicate a technique for selecting non-dominated dual solutions and learn from each iteration of Benders. In the second policy, our objective is to determine a trajectory that expedites the attainment of the final subproblem’s dual solution. This approach is can be applied on a specific (or a specific class of) problem. From among different problems on which this technique has been examined, we report computational experiments on two successful cases of real-world problems. Our results confirm that incorporating these learned policies significantly enhances the efficiency of the solution process, although the first policy often outperforms the second one.
