| This thesis introduces a neural network approach to the dynamic vehicle dispatching problem. Our goal was to develop a system that would emulate the decision capabilities of a human dispatcher through automated or semi-automated means. To this end, a particular application domain was chosen, namely, a courier service company operating in an urban area.; This application is concerned with the real-time allocation of incoming customer requests, each with a pick-up point and a delivery point, to a fleet of vehicles in movement and the estimation of the new planned routes and schedules for the vehicles that customers are assigned to. The objective is to find a good compromise between conflicting goals like minimization of operations cost (e.g., fuel consumption) and maximization of service quality (e.g., meeting the due dates specified by the customers for the pick-up or the delivery).; Since customer requests are not known in advance the dispatching decisions must be made when new requests come in. Moreover, customers that have already been assigned to a vehicle, and are waiting to be serviced, impose certain constraints on the decision. Therefore, the dispatching task is very difficult for a human being, and it takes a lot of experience and judgment to master this skill.; Observing the fact that expert dispatchers can do the job remarkably well, but can hardly explain their decision making, this thesis will explore the avenue of emulating the dispatching expertise through automatic machine learning techniques, more specifically neural network techniques.; This thesis makes the following contributions. First, the dispatching problem is modeled as a multiattribute choice problem. That is, characteristics are associated with each vehicle or driver in regard to the current dispatching situation (e.g., detour to service the new request, pick-up time, etc...). Then, based on this attribute description, a neural network estimates the quality of each vehicle, in order to focus the attention of the dispatcher on the best alternatives. The neural network develops this ability after a training phase on previous decisions taken by the dispatcher. Secondly, an interactive-graphic vehicle dispatching system is developed which provides both "passive" aid (like informative displays on the current routes) and "active" aid through the suggestions of the neural network.; Experiments were performed with data obtained from a courier service company. These experiments include learning simple, systematic decision rules, decisions taken by an expert in vehicle routing (but not a professional dispatcher), as well as decisions taken by a professional dispatcher. In each case, the neural network exhibited good learning capabilities. A comparison between the neural network model and a linear programming model is also provided in the appendix, at the end of the thesis. |
