A new bilevel formulation for the vehicle routing problem and a solution method using a genetic algorithm

  • Authors:

  • Yannis Marinakis;Athanasios Migdalas;Panos M. Pardalos

  • Affiliations:

  • Decision Support Systems Laboratory, Department of Production Engineering and Management, Technical University of Crete, Chania, Greece 73100;Decision Support Systems Laboratory, Department of Production Engineering and Management, Technical University of Crete, Chania, Greece 73100;Department of Industrial and Systems Engineering, University of Florida, Gainesville, USA

  • Venue:
  • Journal of Global Optimization
  • Year:
  • 2007

Quantified Score

Hi-index 0.00

Visualization

Abstract

The Vehicle Routing Problem (VRP) is one of the most well studied problems in operations research, both in real life problems and for scientific research purposes. During the last 50 years a number of different formulations have been proposed, together with an even greater number of algorithms for the solution of the problem. In this paper, the VRP is formulated as a problem of two decision levels. In the first level, the decision maker assigns customers to the vehicles checking the feasibility of the constructed routes (vehicle capacity constraints) and without taking into account the sequence by which the vehicles will visit the customers. In the second level, the decision maker finds the optimal routes of these assignments. The decision maker of the first level, once the cost of each routing has been calculated in the second level, estimates which assignment is the better one to choose. Based on this formulation, a bilevel genetic algorithm is proposed. In the first level of the proposed algorithm, a genetic algorithm is used for calculating the population of the most promising assignments of customers to vehicles. In the second level of the proposed algorithm, a Traveling Salesman Problem (TSP) is solved, independently for each member of the population and for each assignment to vehicles. The algorithm was tested on two sets of benchmark instances and gave very satisfactory results. In both sets of instances the average quality is less than 1%. More specifically in the set with the 14 classic instances proposed by Christofides, the quality is 0.479% and in the second set with the 20 large scale vehicle routing problems, the quality is 0.826%. The algorithm is ranked in the tenth place among the 36 most known and effective algorithms in the literature for the first set of instances and in the sixth place among the 16 algorithms for the second set of instances. The computational time of the algorithm is decreased significantly compared to other heuristic and metaheuristic algorithms due to the fact that the Expanding Neighborhood Search Strategy is used.