Machine learning for satisficing operational decision making: A case study in blood supply chain

Machine learning (ML) has attracted recent attention in solving constrained optimization problems to reduce computational time. In this article, we employ multi-output ML models including light gradient-boosting machine (LGBM) and multilayer perceptron (MLP) to predict the solution to a constrained optimization problem, explore the impacts of selecting different loss functions, and evaluate their performance by looking at the utility of predicted decisions, i.e., the associated cost components and the feasibility of predicted solutions We implement our approach in a blood supply chain where hospitals collaborate on transshipment to meet demand. We use demand distributions fitted to real data to evaluate the performance of the predictive models by analysis of the utility of forecasts, including total cost, inventory holding, transshipment, outdated unit, and shortage costs associated with predicted decisions. The results of our case study show that an LGBM model with the mean absolute deviation loss function provides solutions with only 2% higher total cost than a stochastic optimization model. Compared to an empirical policy, it could reduce transshipment and shortage costs by 23% and 6%, respectively. Therefore, altering the loss function of ML models can provide appropriate solutions to optimization problems, and various loss functions should be probed accordingly.