In this paper, a novel Particle Swarm Optimization (PSO)
approach is applied to solve the CVRP having stochastic demands
with no known distributions. Since PSO solves problems with continuous
variables and the VRP is a combinatorial optimization
problem, a decoding method is needed to apply PSO to solve the
VRP. We have developed a novel decoding method for interpreting
PSO solutions for the VRP. The proposed M1 decoding method
includes three local search operators in a VNS loop in order to significantly
improve the quality of the solutions. It has been tested
against two recent and state-of-the-art decoding methods from
Ai and Kachitvichyanukul (2009b). The results clearly show that
our proposed approach is a superior and preferable PSO method
in some CVRP instances. The whole proposed PSO approach has
also been tested for the Stochastic VRP (SVRP). Results show that
the solutions of this algorithm are applicable in larger scale problems
and resist perturbations to an acceptable extent. Although
the exact robust algorithm gives solutions without any unmet
demands, our proposed method (even with some unmet demands)
has less robustness cost. Also, in all cases, the proposed method has
produced a feasible solution while the other method has not (in
many cases).
Our intention for future research is to actually test the proposed
PSO in a production environment with a real SVRP and to actually
measure unmet vs. robustness costs. A close study of the performance
of the proposed PSO parameters can be the subject of
further researches.