Rollout及其并行求解算法在多类商品库存控制中的应用  

Parallelizing Rollout Algorithms for Multi-product Inventory Control

作　　者：李豹[1] 程文娟[1] 周雷[1] 唐昊[1] 

机构地区：[1]合肥工业大学计算机与信息学院,安徽合肥230009

出　　处：《系统仿真学报》2007年第17期3883-3887,共5页Journal of System Simulation

基　　金：国家自然科学基金项目(60404009);安徽省自然科学基金项目(050420303;070416242);安徽高校自然科学研究重点项目(KJ2007A063)

摘　　要：Rollout算法是Bertsekas提出的求解马尔科夫决策过程(MDP)问题的一种仿真优化算法。文章研究Rollout算法求解多类商品库存控制问题,给出了基于性能势和神经元动态规划的Rollout优化算法。另外,为了降低运算时间,文章提出了两种Rollout并行求解算法,并讨论了这两种并行算法各自的适用场合。实验结果表明,Rollout算法能满足模型未知系统的优化要求,具有较好的并行性能。The rollout algorithm （RA） is a simulation and optimization method, proposed by Bertsekas, for solving Markov decision processes （MDPs）. An extension of the rollout algorithm was derived that was applied to multi-product inventory control. The rollout algorithm was given based on performance potentials and neuro-dynamic programming. In addition, since the rollout algorithm had a very strong inherent parallelism, two methods for parallelizing this algorithm were proposed to reduce the computation time, and their performance was analyzed. Some examples of multi-product inventory control were proposed by using the rollout algorithm. The numerical results show that the rollout algorithm can meet the requirement of the systems with unknown parameters, and has a good parallel performance. Key words： Rollout algorithms, inventory control, Markov decision process, performance potentials, parallel algorithms, neuro-dynarnic programming

关 键 词：ROLLOUT算法 库存控制 MARKOV决策过程 性能势 并行算法 神经元动态规划 

分 类 号：TP202[自动化与计算机技术—检测技术与自动化装置]