Incorporation of parameter prediction models of different fidelity into job shop scheduling

Scheduling of industrial job shop processes is normally conducted using estimates of parameters (e.g. processing times) defining the optimization problem. Inaccuracy in these estimated parameters can significantly affect the optimality, or even feasibility, of the scheduling solution. In this work, we incorporate data-driven parameter prediction models of different fidelity into a unit-specific continuous time scheduling model, and investigate the dependency of the solution quality on the prediction model fidelity. Our high-fidelity prediction model is based on Gaussian processes (GP); more specifically we use the maximum a posteriori probability (MAP) estimate. The low and medium-fidelity prediction models rely on determining the average processing time or average processing rate, respectively, from the dataset. In our test case, involving prediction of taxi durations in New York City, the use of GP prediction model yielded, on average, 5.8% and 1.8% shorter realized make spans in comparison to using the low and medium-fidelity prediction models, respectively.