N. Gómez-Vargas, R. Blanquero, E. Carrizosa
Uncertainty in logistic decisions regarding the multiple parameters that model these problems- demands, travel times, etc.- requires of approaches within Operations Research (OR) that leverage auxiliary data (e.g., congestions or weather) to predict these parameters and then address decision-making. This falls into prescriptive analytics, which aim is at suggesting the best proactive options in order to take advantage of the predicted outcomes. For the multiple-output regression we chose neural networks (NNs) due to their ability of learning a continuous function that naturally captures the complex relationships between inputs and also jointly models dependencies in the outputs. The effectiveness of the prescriptions depends on how well the predictions are integrated with the OR problem. We present a tailored NN model that both accounts for the decision objective in its loss function and adapts to the structure of the problem, avoiding the propagation of unimportant information.
Keywords: multi-output learning, neural networks, prescriptive analytics
Scheduled
GT04 Multivariate Analysis and Classification V. Mathematical Optimization and Data Science
June 8, 2022 12:40 PM
Cloister room
