SBIR PHASE II: A Neuro-Dynamic Programming Approach to Stochastic Control

这个小企业创新研究二期项目将开发神经动态规划（NDP）方法来解决一个商业上重要的复杂随机控制问题——供应链管理问题。第一阶段的研究成功地证明，在作为测试平台的一类供应链上，NDP算法可以比公认的启发式算法显著节省成本。第二阶段的项目将进一步发展和简化NDP方法。此外，算法将被普遍化，以便它们适用于非常广泛的现实供应链，包括制造和分销网络。在这些新问题上的表现将通过广泛的实验和与当前最先进的启发式方法的比较来评估。一旦NDP方法得到充分开发，将寻求许可安排，将基于NDP的优化模块集成到目前广泛使用的许多供应链管理产品中。解决供应链管理物流的改进方法将对所有行业的公司产生巨大的商业利益。将该技术集成到制造业中广泛使用的现有商业供应链管理软件产品中，可以提高许多美国公司的效率。此外，本研究开发的一般NDP方法具有更广泛的潜在范围，因为它们可以用于解决在许多国家重要领域出现的其他复杂随机控制问题，包括过程控制，排队和调度以及数据网络优化。* * * p: / sbir / yhashimi /摘要/ 9704090.文档

*** Lee 9704090 This Small Business Innovation Research Phase II project will develop neuro-dynamic programming (NDP) methods to address a commercially important complex stochastic control problem--that of supply-chain management. The Phase I research successfully established that NDP algorithms could lead to significant savings over well-accepted heuristics on a class of supply chains that was used as a testbed. This Phase II program will further develop and streamline the NDP methodology. Furthermore, algorithms will be generalized so that they apply to a very broad class of realistic supply chains, including manufacturing and distribution networks. Performance on these new problems will be assessed through extensive experimentation and comparison with current state-of-the-art heuristics. Once the NDP methodology is fully developed, licensing arrangements will be sought to integrate NDP-based optimization modules into the many supply-chain management products currently in widespread use. An improved approach for addressing the logistics of supply-chain management will be of great commercial interest to companies across all industries. Integration of the technology into existing commercial supply-chain management softuare products widely used in manufacturing can enhance the efficiency of numerous U.S. corporations. Furthermore, the general NDP methods developed in this research have a much broader potential scope, in that they can be used to address other complex stochastic control problems that arise in many areas of national importance, including process control, queuing and scheduling, and data network optimization. *** p:/sbir/yhashimi/abstracts/9704090.doc