Improvement of PSO algorithm by memory based gradient search - application in inventory management

摘要

Advanced inventory management in complex supply chains requires effective androbust nonlinear optimization due to the stochastic nature of supply and demandvariations. Application of estimated gradients can boost up the convergence ofParticle Swarm Optimization (PSO) algorithm but classical gradient calculationcannot be applied to stochastic and uncertain systems. In these situationsMonte-Carlo (MC) simulation can be applied to determine the gradient. Wedeveloped a memory based algorithm where instead of generating and evaluatingnew simulated samples the stored and shared former function evaluations of theparticles are sampled to estimate the gradients by local weighted least squaresregression. The performance of the resulted regional gradient-based PSO isverified by several benchmark problems and in a complex application examplewhere optimal reorder points of a supply chain are determined.