In today’s world, organizations across various industries face the challenge of efficiently scheduling their production processes and rostering their workforce optimally. However, despite consistent improvements in combinatorial optimization software for scheduling and rostering, the complexity of this task continues to grow due to uncertainty about multiple factors such as employee availability, demand fluctuations, supplier variability, variable prices, the impact of weather and the increasing need for energy efficiency. Machine learning can be used to make estimates about these uncertain factors, but the real challenge is in integrating predictions and the optimization of scheduling and rostering problems. Or more precisely *that predictions and optimization over these predictions need to be developed and evaluated together*.
While many combinatorial optimisation solvers for solving scheduling and rostering exists, including Constraint Programming and Mixed Integer Programming solvers; few of these solvers can be easily integrated with machine learning libraries. Futhermore, in a machine learning pipeline, the requirements for the solver change. What is needed is a framework for solving prediction + optimization problems that bridges the machine learning and combinatorial optimization solving tools. It should allow actors to discover what a data-driven approach can signifigy to their scheduling and rostering problem, by allowing them to easily experiment and prototype, both on the learning side, the solving side and the combination of the two.
In my ERC Consolidator project 'Conversational Human-Aware Technology for Optimisation', we started building such a library: CPMpy. We notice an increasing industrial interest in solving Prediction + Optimisation problems, but a lack of unified tools to do so. This project set out to increase the Technological Readiness Level of CPMpy from TRL 4 to 6; and to demonstrate its potential and align it with industry needs.
