Constraint Learning for Operations Research

The field of operations research helps in decision making processes. A particular problem often encountered by the experts in this field is to find optimal decisions while restricted by some constraints. For instance, consider making a schedule for nurses in a hospital, where the goal might be to minimise the total number of working hours for each nurse, while making sure that the schedule satisfies the minimal requirement of nurses in each shift. Traditionally, such problems are solved by first encoding the problem as a formal mathematical model, and then using a general purpose solver to find a solution. These mathematical models are declarative in nature and thus a problem can be modeled without considering how the solver is going to find a solution. Moreover, the general purpose solvers can solve any model in a model class and thus models can be modified without making any changes in the solver. In the past couple of decades the research community has seen a lot of work on designing efficient solvers, making mathematical models a significant part of decision making processes. However, these models are still manually built by the experts, which is not an easy task and requires investment of both time and money. The focus of this thesis is to automatically learn mathematical models from past examples of working and non-working solutions of the model. Learning models from data allows users that are not familiar with modelling techniques to make use of powerful solvers. Moreover, as the learned models are not black box, an expert user can modify and extend the learned models.

Jaar van publicatie:2021

Toegankelijkheid:Open