Tuning Bayesian Optimization for Problems with Dynamic Resource Constraints

闭环优化处理通过进行实验（例如物理或生化实验）来评估候选解决方案的问题。虽然这种形式的优化在科学界越来越受欢迎，但它可能会受到未经探索的资源问题的影响，因为任何实验都可能需要资源才能进行。在这个博士项目中，我们关注的是了解贝叶斯优化如何受到动态资源约束的影响-一种建模资源可用性的约束（例如，原材料、存储、计算能力和存储、熟练工程师、设备和机器、预算，等）进行物理实验或运行耗时的模拟所需的搜索策略，以及解决这一特定问题的搜索策略的开发。我们希望该项目能做出一些贡献：1扩展动态资源约束的文献，包括其他类型的约束和具体的工业实例。2从经验和理论上进行调查，动态资源约束对贝叶斯优化的影响（与迄今为止文献中主要采用的运筹学方法相反）。3开发和分析在贝叶斯优化器上增强的各种搜索策略，以应对动态资源约束。4解决所开发方法的超参数优化问题。5用可解释的AI增强方法，以提供最终的用户理解影响资源使用和优化性能的关键特征，并且潜在地利用来自解释的信息来通知优化过程。噪声、不确定性、鲁棒性和非均匀实验成本的存在也将酌情纳入分析。搜索策略的灵感将来自许多研究领域，包括机器学习，统计学和运筹学。由于贝叶斯优化是一种机器学习技术，这直接符合EPSRC对人工智能技术的战略重点。此外，这与EPSRC对人工智能技术的C1目标非常一致，即“实现有竞争力的数据驱动型经济”，因为该工作旨在开发智能工具来解决现实世界中的动态工业优化问题，这些问题可能对各行各业的各种挑战产生重大影响。

Closed-loop optimization deals with problems in which candidate solutions are evaluated by conducting experiments, e.g. physical or biochemical experiments. Although this form of optimization is becoming more popular across the sciences, it may be subject to rather unexplored resourcing issues, as any experiment may require resources in order to be conducted. In this PhD project we are concerned with understanding how Bayesian optimization is affected by dynamic resource constraints - a type of constraint that models the availability of resources (e.g. raw materials, storage, computing power and storage, skilled engineers, equipment and machines, budget, etc) required to conduct a physical experiment or run a time-consuming simulation - and the development of search strategies to tackle this particular problem issue. We expect the project to make a number of contributions:1 Expand on the literature of dynamic resource constraints with additional types of constraints and concrete industrial examples.2 Investigate, empirically and theoretically, the effect of dynamic resource constraints on Bayesian optimization (as opposed to operations research methods as done largely in the literature so far).3 Develop and analyze various search strategies augmented on a Bayesian optimizer for coping with dynamic resource constraints.4 Address the issue of hyperparameter optimization for the methods developed.5 Augment the methodology with explainable AI to provide end-users with an understanding of key features affecting resource usage and optimization performance, and potentially utilise information from the explanations for informing the optimization process. The presence of noise, uncertainty, robustness and non-homogenous experimental costs will also be factored into the analysis as deemed appropriate. Inspiration for search strategies will be drawn from a number of research fields including Machine Learning, Statistics, and Operations Research. As Bayesian optimisation is a Machine Learning technique, this directly aligns with the EPSRC's strategic focus on Artificial Intelligence technologies. Moreover, this aligns strongly with the EPSRC's C1 ambition for AI technologies, to "Enable a competitive, data-driven economy", as the work looks to develop smart tools to solve real-world dynamic industrial optimisation problems that can have significant impact on a broad variety of challenges across a range of industries.