The Field Technician Scheduling Problem: A Predict-then-Optimise Approach Using Machine Learning and Constraint Programming

The Field Technician Scheduling Problem (FTSP) in Field Service Management (FSM) requires assigning skilled technicians to geographically dispersed tasks under strict Service Level Agreements (SLAs) and finite capacity. This thesis investigates the predict-then-optimise paradigm by developing a decision support prototype that bridges data-driven duration estimation and exact Constraint Programming. A Random Forest model, trained on 3,000 synthetic historical assignments, predicts job durations with a cross-validated coefficient of determination ($R^2$) of 0.748, correcting a systematic duration bias: actual historical durations exceed planned durations by 14.8\% on average. Benchmarking shows that most of this gain comes from removing that bias rather than from model complexity: linear regression scores marginally better on aggregate error, and the tree-based ensemble is retained as the implemented predictor, with feature-importance summaries providing operational inspection. These predicted durations drive a Google OR-Tools CP-SAT model that schedules technicians subject to skill compatibility and daily capacity limits, with multi-day reservations and penalties for unfilled crew slots. The integration yields a critical diagnostic result: under a formulation requiring each job to fit within a single technician-day, the modelled workload renders the schedule largely infeasible, with 762 jobs exceeding the daily limit. A four-lever capacity strategy incorporating multi-day rolling carryover, selective overtime, targeted workforce augmentation, and priority-weighted penalties recovers coverage to 851 jobs under ML inputs and 904 jobs under static inputs, with jobs receiving no technician assignment confined to normal priority. To maintain computational tractability, the 30-day horizon is decomposed into sequential daily scheduling windows rather than solved as a monolithic global optimisation, and travel is modelled as a fixed cost from each technician base to the job site without optimising inter-job routes. Within this bounded scope, the framework enables interactive what-if scenario analysis, allowing management to stress-test capacity assumptions and shift from reactive dispatching toward proactive workforce planning.