Forschungsthemen
[MA] Transfer Learning for Multi-surrogate-model Optimization
Surrogate-model-based optimization is widely used, when the evaluation of a real target system is expensive. However, when the optimization budget is limited to a single or several evaluations, transfer learning, i.e., the usage of data gathered in previously performed optimization experiments, becomes the only feasible option. Recent work in surrogate-model-based optimization showed that multi-model optimization can be extremely efficient in complex search spaces. However, state-of-the-art approaches for transfer learning consider a single-model setup only. The goal of this thesis is to improve the quality delivered by a multi-surrogate-model optimization approach within a limited budget. The research objective is to develop a strategy of applying transfer learning to a multi-model setup.
For this thesis the following tasks have to be fulfilled:
* Literature analysis covering closely related work.
* Development of a strategy for multi-model transfer learning.
* Implementation of the developed strategy.
* Evaluation of the developed approach using a completion-algorithm-based target system.
* (Optional) Evaluation of the developed approach using an anytime-algorithm-based target system.
Betreuer: Dmytro Pukhkaiev