Speaker
Description
Efficient dispatching of timber trucks is a major determinant of cost, productivity, and supply reliability in forest transport systems. In operational contexts, where transport distances are long, pickup locations vary daily, and mills operate under strict time windows, dispatching decisions must account for dynamic constraints including queue conditions, service times, driver duty limits, and interactions among trucks competing for shared resources. Traditional truck scheduling models can optimize fleet plans for a full day, but their computational requirements limit their use for real-time decision support when conditions change during operations.
This study presents a hybrid framework that integrates optimization and machine learning (ML) to enable rapid, adaptive dispatching decisions. A fleet-level truck scheduling model is first used as a “teacher” to generate optimized daily schedules across multiple simulated scenarios representing realistic Australian forest transport conditions. These solutions are decomposed into sequential decision points, capturing system state variables such as truck locations, remaining duty time, active transport tasks, origin–destination distances and costs, queue estimates at pickup points and mills, and mill time-window constraints. These data are used to train an ML model that learns to predict the next best transport task for each truck.
Because mill destinations remain relatively stable while pickup locations change daily, the model is structured to generalize across new harvesting configurations. Results indicate that the learned dispatching policy can approximate optimization-quality decisions while operating fast enough for real-time deployment. The proposed approach demonstrates how combining optimization-based planning with data-driven learning can support intelligent, responsive dispatching systems for modern forest transport operations.
| Keywords | dispatching; machine learning; optimization; |
|---|
