Supervised Learning of the Optimal Objective Function Value in Chemical Production Scheduling

Mixed-integer programming (MIP) can be used to formulate and solve complex production scheduling problems in the field of process systems engineering. However, the solution of MIP models may require a long computing time due to the combinatorial complexity of the problems. In this work, we propose supervised learning models to predict the optimal objective function value on four classes of scheduling problems, which can be useful in a number of settings. To improve the accuracy of the prediction models, we device a number of machine learning features based on the instance parameters. The studied objective functions are cost and makespan minimization. Based on the results, the prediction accuracy is high─the coefficients of determination with the best prediction models are r² > 0.97 on the four classes of problems. These predictions allow us to predict how different problem features (e.g., new orders or disturbances) affect the optimal objective function value.