Collaborative Research: A Framework for Effective Optimization via Simulation

我们将开发算法（和支持理论）来优化相对于离散决策变量的随机系统的预期性能。我们假设感兴趣的随机系统由仿真模型表示，因此该系统的性能只能用噪声来估计。我们的重点是不利用特定问题结构的“通用”优化技术，因为我们希望我们的技术适合包含在通用仿真软件中。目标是生成具有可证明的渐近性能、有竞争力的有限时间性能和有效的终止统计推断的算法。我们方法的关键是：(1)我们的算法将在保证渐近收敛的全局指导框架内工作，同时给予我们广泛的自由度来积极和自适应；(2)在此框架内，我们将推行积极的地方改善计划；(3)对局部改进方案进行改进，使其具有高效的选择误差控制，确保在存在估计误差的情况下也能改进；(4)我们将在算法终止时提供有效的统计推断，以便报告为最佳的解决方案将是搜索实际访问的所有解决方案中最好的或接近最好的，具有预先指定的置信度。在美国，计算机模拟被广泛用于设计和改进（“优化”）受不确定性影响的制造业、服务业、军事、电信和金融系统。我们的研究将提供理论上合理的优化算法，可以纳入新的或现有的仿真软件包。这项研究非常需要，因为每天模拟用户都在使用商业产品制定并试图解决通过模拟进行优化的问题，而这些商业产品忽略或只是稍微注意到模拟实验包含不确定性。这些商业产品通常工作得很好，但它们也可能被严重误导，用户没有任何指示或保护，可能导致错误和昂贵的决定。几乎所有商业仿真建模软件包中都有可用的优化工具，这意味着通过仿真进行优化的问题将得到“解决”。问题是，这些问题能否用理论上合理的算法有效地解决，这些算法可以为它们的性能提供具体的保证和推断。我们的研究目标是开发这样的优化通过模拟算法，代表了在理论和实践的艺术状态的实质性进展。

We will develop algorithms (and supporting theory) for optimizing the expected performance of a stochastic system with respect to discrete decision variables. We assume that the stochastic system of interest is represented by a simulation model, and hence that the performance of this system can only be estimated with noise. Our focus is on ``general-purpose'' optimization techniques that do not exploit particular problem structure, because we want our techniques to be suitable for inclusion in general-purpose simulation software. The goal is to produce algorithms that have provable asymptotic performance, competitive finite-time performance, and valid statistical inference at termination. The keys to our approach are (1) our algorithms will work within a global guidance framework that guarantees asymptotic convergence, while giving us wide latitude to be aggressive and adaptive; (2) within this framework, we will embed aggressive local-improvement schemes; (3) we will enhance the local-improvement schemes with highly efficient selection-error control to insure improvement even in the presence of estimation error; and (4) we will provide valid statistical inference at algorithm termination so that the solution reported as best will be the best, or near best, of all those solutions actually visited by the search, with a prespecified confidence level.In the United States, computer simulation is widely used to design and improve ("optimize") manufacturing, service, military, telecommunication and financial systems that are subject to uncertainty. Our research will provide theoretically sound optimization algorithms that can be incorporated into new or existing simulation software packages. There is a critical need for this research, because every day simulation users are formulating and attempting to solve optimization-via-simulation problems using commercial products that ignore, or only slightly notice, that the simulation experiment incorporates uncertainty. These commercial products often work well, but they can also be dramatically misled, and the user has no indication of, or protection against, the incorrect and costly decisions that may result. The availability of optimization tools in nearly all commercial simulation modeling packages implies that optimization-via-simulation problems will be "solved." The question is whether they will be solved efficiently with theoretically sound algorithms that provide specific guarantees of, and inference on, their performance. The goal of our research is to develop such optimization-via-simulation algorithms, representing a substantial advance over the state of the art in both theory and practice.