Algorithm portfolios for logistics optimization considering stochastic demands and mobility allowance

The vehicle routing problem with stochastic demand (VRPSD) is a well known NP‐hard problem. The uncharacteristic behaviour associated with the problem enhances the computational efforts required to obtain a feasible and near‐optimal solution. This paper proposes an algorithm portfolio methodology based on evolutionary algorithms, which takes into account the stochastic nature of customer demand to solve this computationally complex problem. These problems are well known to have computationally complex objective functions, which make their solutions hard to find, particularly when problem instances of large dimensions are considered. Of particular importance in such situations is the timeliness of the solution. For example, Apple was forced to delay their shipments of iPads internationally due to unprecedented demand and issues with their delivery systems in Samsung Electronics and Seiko Epson. Such examples illustrate the importance of stochastic customer demands and the timing of delivery. Moreover, most of the evolutionary algorithms, known for providing computationally efficient solutions, are unable to always provide optimal or near optimal solutions to all the VRPSD instances within allocated time interval. This is due to the characteristic variations in the computational time taken by evolutionary algorithms for same or varying size of the VRPSD instances. Therefore, this paper presents portfolios of different evolutionary algorithms to reduce the computational time taken to resolve the VRPSD. Moreover, an innovative concept of the mobility allowance (MA) in landmoves based on the levy’s distribution function has been introduced to cope with real situations existing in vehicle routing problems. The proposed portfolio approach has been evaluated for the varying instances of the VRPSD. Four of the existing metaheuristics including Genetic Algorithm (GA), Simulated Annealing (SA), Artificial Immune System (AIS), TABU Search (TS) along with new neighbourhood search, are incorporated in the portfolios. Experiments have been performed on varying dimensions of the VRPSD instances to validate the different properties of the algorithm portfolio. An illustrative example is presented to show that the set of metaheuristics allocated to certain number of processors (i.e. algorithm portfolio) performed better than their individual metaheuristics.