| Continuous vehicle dispatching is an NP-complete problem, thus requiring local methods to solve the problem. This paper presents a multi-stage approach to solving the continuous vehicle dispatching problem. In the first stage, linear programming is used to create an abstract schedule. In the second stage, a Monte Carlo reinforcement learning algorithm is used to operationalize the schedule, achieving near optimal results. The reinforcement learning algorithm uses the abstract schedule as its state representation. Tests are performed to compare this method against another multi-stage approach using linear programming and a load balancing local search method. The reinforcement learning approach is shown to perform better than the load balancing local search method. |
