Designing actively robust products with Bayesian Optimisation

研究背景在某些工程背景下，如风力涡轮机的设计，重要的是设计人员要对环境的变化保持鲁棒性。术语“主动鲁棒性”描述了具有一定能力以适应其环境的设计，例如通过改变风力涡轮机叶片的桨距。优化这样的设计是一个困难的问题，因为在具有特定参数设置的特定环境中评估单个设计通常需要运行昂贵的仿真。贝叶斯优化是一种优化技术，旨在尽可能提高样本效率，使其成为研究快速设计主动稳健产品的完美目标。优化问题广泛存在于工程，机器学习和数学建模中。在某些情况下，查询目标函数是非常昂贵的，最常见的是因为它需要很长的时间来评估。例如，问题可能是优化风力涡轮机叶片的形状，为了评估候选设计，我们需要运行缓慢的计算流体动力学模拟。然而，还存在许多其他应用，从调整神经网络中的超参数到为制药过程推荐湿实验室实验。贝叶斯优化是一种全局优化方法，它在优化过程中尽可能提高样本效率，工程设计中进一步考虑的是鲁棒性。产品必须部署到条件未知和/或可能发生变化的环境中。在风力涡轮机的情况下，风速和风向可以全天变化并且可以是阵风。而不是建立一个静态的设计，以适应所有的环境，'主动鲁棒性'是指一个设计的能力，以作出反应，在它的环境中的变化。在这个项目中，我们建议应用贝叶斯优化方法，以加快寻找最佳设计的过程中，可以作出反应，以目前的anvirenmental条件。在一个标准的鲁棒贝叶斯优化问题中，我们寻求在一系列环境条件下优化设计的性能，例如，通过对这些条件的期望或考虑最坏情况的结果。研究的目的和目标1.开发贝叶斯优化方法，以帮助和加快积极稳健产品的设计。2.在标准基准问题和真实的上测试该方法。世界设计问题。演示的方法，以现实世界的设计问题所提供的外部partner.The新颖的研究methodologyBayesian优化的新奇性还没有被应用到主动鲁棒性的问题。在博士课程中，我们将开发新的替代模型、获取函数和优化策略，以选择尽可能多的实验，从而揭示最佳主动稳健设计的信息。潜在的影响、应用和益处外部合作伙伴GE可能会将我们的方法集成到其内部优化工具包中。通用电气的许多团队都在使用这一方法，因此我们的研究有可能影响许多项目。研究与汇款的关系研究福尔斯属于工程设计和人工智能技术的EPSRC主题，根据所采取的方向，与控制工程和运筹学有潜在的联系。通过GE在电力系统中的应用也将该项目与能源和风力发电的主题联系起来。

The context of the researchIn certain engineering contexts, such as the design of wind turbines, it is important for designes to be robust to changes in their environment. The term 'active robustness' describes designs which have some capacity to adapt to their environment, for example by changing pitch of a wind turbine blade. Optimising such a design is a difficult problem becuase evaluating a single design in a particualr environment with specific parameter settings typically requires running an expensive simulation. Bayesian optimistation is an optimisation technique aimed at being as sample efficient as possible, making it a perfect target for research into faster design of actively robust products.Optimisation problems occur extensively thorughout engineering, machine learning and mathematical modelling. In some cases, querying the objective function is very expensive, most commonly becuase it takes a long time to evaluate. For example, the problem might be to optimize the shape of a wind turbine blade, where in order to evaluate a candidate design we need to run a slow computational fluid dynamics simulation. Many other applications exist however, ranging from tuning hyper-parameters in a neural network to recommending wet lab experiements for a pharmaceutical process. Bayesian optimisation is a global optimisation method which attempts to be as a sample efficient as possible duing the optmisation process.A further consideration in engineering design is that of the robustness. Products must be deployed into environments where the conditions are unknown and/or subject to change. Iin the case of a wind turbine, the wind speed and direction can change throughout the day and can be gusty. rather than building a static design to suit all environments, 'active robustness' refers to the ability of a design to react to changes in it's environment.In this project, we propose to apply Bayesian optimisation methods to speed up the process of finding optimal designs which can react to their current anvirenmental conditions. In a standard robust Bayesain optimisation problem we seek to optimise the performance of a design over a range of environmental conditions, for example by taking expectation over these conditions or considering a worst case outcome.The aims and objectives of the research1.Develop Bayesian optimisation methodology to aid and speed up the design of actively robust products.2.test the methodology on standard benchmark problems and real-world design problems.3. Demonstrate the applicability of the methodology to real-world design problems provided by the external partner.The novelty of the research methodologyBayesian optimisation has not yet been applied to the problem of active robustness. Over the course of the PhD we will develop novel surrogate models, acquisition functions and optimisation strategies to select experiments which reveal as much information as possible ablout the optimal actively robust design.The potential impact, applications and beneftsIt is possible that the xternal partner, GE, will integrate our methodology into their in-house optimisation toolkit. This is used by many teams across GE, so our research has the potential to impact many projects.How the research relates to the remitThe research falls into the EPSRC themes of engineering design and artificial intelligence technologies, with potential links to control engineering and operations research depending on the direction taken. The application to power systmes through GE also relates the project to the themes of energy and wind power.