Multi-Label Classification For The Generation Of Sub-Problems In Time-Constrained Combinatorial Optimization

This paper addresses the resolution of combinatorial optimization problems presenting some kind of recurrent structure, coupled with machine learning techniques. Stemming from the assumption that such recurrent problems are the realization of an unknown generative probabilistic model, data is collected from previous resolutions of such problems and used to train a supervised learning model for multi-label classification. This model is exploited to predict a subset of decision variables to be set heuristically to a certain reference value, thus becoming fixed parameters in the original problem. The remaining variables then form a smaller subproblem whose solution, while not guaranteed to be optimal for the original problem, can be obtained faster, offering an advantageous tool for tackling time-sensitive tasks.