Extending model-based optimization with resource-aware parallelization and for dynamic optimization problems

Abstract

This thesis contains two works on the topic of sequential model-based optimization (MBO). In the first part an extension of MBO towards resource-aware parallelization is presented and in the second part MBO is adapted to optimize dynamic optimization problems. Before the newly developed methods are introduced the reader is given a detailed introduction into various aspects of MBO and related work. This covers thoughts on the choice of the initial design, the surrogate model, the acquisition functions, and the final optimization result. As most methods in this thesis rely on the Gaussian process regression it is covered in detail as well. The chapter on “Parallel MBO” dives into the topic of making use of multiple workers that can evaluate the black-box and especially focuses on the problem of heterogeneous runtimes. Strategies that tackle this problem can be divided into synchronous and asynchronous methods. Instead of proposing one configuration in an iterative fashion, as done by ordinary MBO, synchronous methods usually propose as many configurations as there are workers available. Previously proposed synchronous methods neglect the problem of heterogeneous runtimes which causes idling, when evaluations end at different times. This work presents current methods for parallel MBO that cover synchronous and asynchronous methods and presents the newly proposed Resource-Aware Model-based Optimization (RAMBO) Framework. This work shows that synchronous and asynchronous methods each have their advantages and disadvantages and that RAMBO can outperform common synchronous MBO methods if the runtime is predictable but still obtains comparable results in the worst case. The chapter on “MBO with Concept Drift” (MBO-CD) explains the adaptions that have been developed to allow optimization of black-box functions that change systematically over time. Two approaches are explained on how MBO can be taught to handle black-box functions where the relation between input and output changes over time, i.e. where a concept drift occurs. The window approach trains the surrogate only on the most recent observations. The time-as-covariate approach includes the time as an additional input variable in the surrogate, giving it the ability to learn the effect of the time. For the latter, a special acquisition function, the temporal expected improvement, is proposed.