Optimization and Performance Evaluation of Network Models

网络模型的优化和性能评估

• 批准号: 9972957
• 负责人: Sean Meyn
• 金额: $ 23.36万
• 依托单位: University of Illinois at Urbana-Champaign
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 1999
• 资助国家: 美国
• 起止时间: 1999-08-15 至 2002-12-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=9972957&HistoricalAwards=false
• 关键词: Optimization; Performance; Evaluation; Network; Models

ECS-9972957MeynThe proposed research will consider the development of recent approaches to optimal control, reinforceraent learning, and performance evaluation for complex systems. The methods to be employed are based upon recent work conducted by the PI on Markov chain stability theory, linear programming methods, and fluid model approximation.Research on dynamic optimization will be based on a recent discovery of the PI which establishes a relationship between the optimal control of a stochastic model, and the optimal control of a simpler 'leaky bucket' fluid model. The latter gives rise to a constrained L1 optimal control problem which can be solved in many practical examples.The optimization theory will be improved through a deeper analysis of linear programming methods for approximating the value function in the optimal control problem of interest.Reinforcement learning algorithms will be considered for policy improvement in stochastic models. The PI will again base the analysis on fluid model approximations.Specific applications include the development of scheduling and routing algorithms for multiclass queueing networks. The algorithms will be based upon translations of fluid policies; tranlations of fluid value function approximations; and the application of reinforcement learning techniques.

ECS-9972957Meyn 拟议的研究将考虑复杂系统的最优控制、强化学习和性能评估的最新方法的发展。 所采用的方法基于 PI 最近在马尔可夫链稳定性理论、线性规划方法和流体模型近似方面进行的工作。动态优化的研究将基于 PI 的最新发现，该发现建立了最优值之间的关系。随机模型的控制，以及更简单的“漏桶”流体模型的最优控制。 后者产生了一个有约束的 L1 最优控制问题，可以在许多实际例子中解决。通过更深入地分析线性规划方法来逼近感兴趣的最优控制问题中的值函数，将改进优化理论。强化学习算法将考虑对随机模型进行政策改进。 PI 将再次基于流体模型近似进行分析。具体应用包括开发多类排队网络的调度和路由算法。 该算法将基于流动策略的转换；流体值函数近似的翻译；以及强化学习技术的应用。