Neighborhood Evaluation on GPU for the DVRP

For many applications of vehicle routing, there is still a large gap between the requirements and the performance of today’s decision support systems. Although there has been a tremendous increase in the ability to solve ever more complex VRPs (partly due to methodological improvements, partly due to the general increase in computing power), the ability to consistently provide better routing plans in shorter time across a variety of instances will give substantial additional savings. In a generic vehicle routing tool, some form of approximate solution method is required [7,8]. Metaheuristics constitute a popular basis for solving rich and large-size VRPs [2,3,6]. For trajectory based metaheuristics, the typical computational bottleneck is neighborhood evaluation. Similarly, the evaluation of individuals is the bottleneck of population based metaheuristics. These tasks are embarrassingly parallel and prime candidates for parallelization [1].