Data-based Approach to Predict Feasibility and Computational Requirement for Chemical Production Scheduling

Abstract

Online scheduling requires frequent re-optimization to generate a schedule repeatedly accounting for updated information. However, if the time between re-optimizations is too short, then finding good, and in some cases even feasible, solutions can become challenging. This work proposed an approach, based on supervised learning techniques, to predict whether a given instance is feasible and, given that it is feasible, what is the computational requirement to solve the instance. Towards this goal, we introduce various types of features related to problem size, scheduling horizon, and processing times and costs that can be derived based on domain knowledge. Logistic regression and random forests models are trained as feasibility classifier and computational time regressor, respectively, using the dataset obtained from a wide variety of instances. Both show good predictive performances: F1 score ∼0.90 and AUC ∼0.98 for the feasibility classification and MSE ∼0.5 for the computational time prediction. Finally, we discuss the features that are shown to be significant in the cases of makespan minimization and cost minimization.