Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.618943
Title: Efficient sequential sampling for global optimization in static and dynamic environments
Author: Morales Enciso, Sergio
ISNI:       0000 0004 5355 9238
Awarding Body: University of Warwick
Current Institution: University of Warwick
Date of Award: 2014
Availability of Full Text:
Access through EThOS:
Access through Institution:
Abstract:
Many optimization problems involve acquiring information about the underlying process to be optimized in order to identify promising solutions. Moreover, in some cases obtaining this information can be expensive, which calls for a method capable of predicting promising solutions so that the global optimum can be found with as few function evaluations as possible. Another kind of optimization problem arises when dealing with objective functions that change over time, which requires tracking of the global optima over time. However, tracking usually has to be quick, which excludes re-optimization from scratch every time the problem changes. Instead, it is important to make good use of the history of the search even after the environment has changed. This thesis revolves around the topic of response surface based sequential sampling for global optimization of expensive-to-evaluate black-box functions under static and dynamic scenarios. Regarding the former scenario, it addresses the high computational cost inherent to Efficient Global Optimization (EGO), a global search algorithm that is known to work well for expensive black-box optimization problems where only few function evaluations are possible, and which uses surrogate models of the fitness landscape for deciding where to sample next. The proposed variant is based on partitioning the space and building local models to accelerate the selection of future sampling locations with a minimal impact on the optimization performance. The linear computational complexity as a function of the number of observations of this extension is shown, and its performance benchmarked to both the original algorithm it extends, and state of the art algorithms. For the latter scenario, we propose and compare four methods of incorporating old and recent information in the surrogate models of EGO in order to accelerate the search for the global optima in a dynamically changing environment. As we demonstrate, exploiting old information as much as possible significantly improves the tracking behavior of the algorithm.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.618943  DOI: Not available
Keywords: QA Mathematics
Share: