Vehicle Routing with Driver Learning for Real World CEP Problems Marcel Kunkel PickPoint AG Ludwig-Eckes-Allee 6 55268 Nieder-Olm, Germany mk@pickpoint.de Michael Schwind Chair of IT-based Logistics Gr¨ uneburgplatz 1 60323 Frankfurt Main, Germany schwind@wiwi.uni-frankfurt.de Abstract—Despite the fact that the vehicle routing problem (VRP) with its variants has been widely explored in operations research, there is very little published research on the VRP concerning real world constraint combinations and large problem sizes. In this work a heuristic solution approach for the VRP with real world constraints is presented driven by the requirements defined by clients in the courier, express and parcel (CEP) delivery industry in order to support their routing plan decisions and driver assignments. The solution algorithm used combines several local-search-based heuristics with constructive elements to solve the VRP with driver learning (VRPDL). As conceptual proof large instances for the capacitated VRP (CVRP) including 560 to 1200 customers are tested and compared to known benchmark results. From those instances new sub-instances are created and sequentially tested adding the driver learning constraint. Finally, the solver is applied to real world CEP instances with driver learning. I. I NTRODUCTION Since its introduction by Dantzig and Ramser [7] the vehicle routing problem (VRP) has developed to one of the most studied problems in operations research. The basic VRP tries to find the shortest route for servicing a certain number of customers while using as less as possible vehicles in the service fleet. VRP problems are combinatorial optimization problems which are NP -hard [20]. Adding further constraints or increasing the problem size of such makes them even more complex and time consuming to solve and require heuristics find a satisfying solution in a reasonable time span. Solution approaches can be grouped into population-based heuristics such as genetic algorithms (GA) [2] or ant colony optimizations (ANT) [8] where what is learned from one or several initial route plans is used to create new route plans and into local-search-based approaches like TABU search [16] or variable neighborhood search (VNS) [19] where a new solution is always generated out of the neighborhood of the current solution. Local search heuristics use several moves techniques in order to explore the neighborhood of a given solution changing one or a few edges or switching the position of a node within or between routes. All local-search-based approaches need intensification and diversification elements in order to deeply explore and diversify from a given solution to find near optimal solutions after a relevant number of iterations. The local search algorithm proposed here combines exchange and 2-opt moves with new targeted move patterns where edge structures are chosen to be changed in order to fulfill the optimization goals. It also uses nearest neighbor and expulsion procedures for speed improvement as well as intensification purposes. This work is influenced by real world problems coming from the courier, express and parcel (CEP) delivery industry, where several thousand customers have to be visited from each depot during one workday. The customer demands (quan- tity requirement) have to be fulfilled without exceeding the truckload of the vehicle (capacity constraint). Drivers develop a certain routine driving in similar areas and visiting the same customers on their routes regularly. This fact will be integrated as customer-based driver learning depending on the regularity with which a certain driver delivers to that customer or the surrounding area. This promises significant savings of time due to better orientation in delivery process, etc. Furthermore, a driver can only start his route in the morning if the sorting of the parcels is done and the vehicle is loaded, which is interpreted as a maximum route length constraint. Solving real world problems with good results means to assign knowledgeable drivers on improved route plans and therefore helping the CEP organization to increase their efficiency in their operations. The remainder of the article is structured as follows. After a short characterization of the VRP with driver learning [23] in the parcel delivery domain a short introduction to recent research into relevant VRP literature and a mathematical formulation is given for the VRPDL. The second chapter also gives insight into the local-search-based solution algo- rithm by presenting the corresponding pseudo-code together with the underlying ideas. In the third chapter comparisons with standard VRP instances introduced by Li, Golden and Wasil [18] are drawn. From those instances new sub-instances are randomly generated and sequentially tested using the our heuristics together with driver learning. Ultimately the setup is used for the time sequential optimization of real world CEP instances with driver learning. II. VEHICLE ROUTING PROBLEM WITH DRIVER LEARNING In this chapter some related literature concerning large and rich VRP jointly with driver learning is discussed. Later the 2012 45th Hawaii International Conference on System Sciences 978-0-7695-4525-7/12 $26.00 © 2012 IEEE DOI 10.1109/HICSS.2012.633 1315