Abstract
We propose an imitation learning framework to enhance the Benders decomposition method. While we aim at generalizing the notion of Alternative Objective Function (AOF) for the subproblem proposed in recently via learning policies to separate cuts (separation subproblem) addressing another issue frequently observed in Benders subproblems —degeneracy. We propose two policies for this purpose each of which is learned on instances of a specific problem. In the first one, we replicate the a technique in selecting non-dominated dual solutions and learn from each iteration of training data. In the second policy, our objective is to determine a trajectory that expedites the attainment of the final subproblem’s dual solution. From among different problem on which this technique has been tested, we report computational experiments on two success cases of real-world problems to train and evaluate these two policies. Our results confirm that incorporating these learned policies significantly enhances the efficiency of the solution process.
Speaker Intro
Shahin Gelareh (Dr. habil.) is an Associate Professor in Operations Research and Logistics at the Université d'Artois in France and a member of the editorial board of Transportation Research Part E. Shahin's research focuses on combinatorial optimization problems arising in logistics (land, maritime, etc.) from the perspective of mathematical programming, algorithmics, and polyhedral. With the recent trend of success in ML-based techniques, Shahin also aims to leverage the power of ML to improve the computational performance of integer programming techniques. Shahin has previously worked at the National University of Singapore, Technical University of Denmark, Polytechnique de Lille, and Portsmouth Business School in the UK.
