An adaptive variable neighbourhood search approach for the dynamic vehicle routing problem

In the traditional vehicle routing problem (VRP), a route plan is pre-determined and remains unchanged afterwards. However in practice, several unforeseen events could occur at any point, which cause traffic congestion and delay to the original planned routes. It is therefore important to re-optimise the routes by taking into consideration the real-time information, leading to the Dynamic VRP (DVRP). While most of the DVRP literature mainly focused on the customer requests as the dynamic aspect, this paper, however, concentrates on the dynamic traffic information based on the level of urgency of the accidents. Critical nodes are introduced into the network to provide a diversion opportunity for the en-route vehicle. This novel concept of 'criticality' is also more practical than the commonly adopted strategy that allows instantaneous diversion at the current vehicle location. We proposed an adaptive variable neighbourhood search (AVNS) algorithm to generate routes in the static environment which is then adapted accordingly for the dynamic setting. This is a two stage VNS approach with the first one acts as a learning stage whose information is then used in stage 2. Here, a smaller number of neighbourhoods and local searches are chosen at each iteration while adopting a pseudo-random selection procedure derived from stage 1. To provide solution diversity, a large neighbourhood search is also embedded into the search. The flexibility and adaptability of our AVNS approach are demonstrated by the high quality solutions obtained when tested on the commonly used VRP datasets, ranging in size from 50 to 1200 customers, which are modified accordingly. In addition, managerial insights related to the tightness of the routes are also presented and analysed.