In contrast to selecting the best algorithmic variant for a given optimization problem statically, i.e. using the same optimizer for the entire optimization process, one could also attempt to adaptively change the structure of the optimizer online. This is known as Adaptive Operator Selection (AOS) or Dynamic Algorithm Configuration, and it builds on the notion that the nature of the search landscape can change during optimization. This means that for some stages of the optimization procedure, a given optimizer might show good performance, while in another stage it might perform poorly. Through AOS, the structure of the optimization algorithm could then be changed to the variant which works best at any given stage of the optimization process. AOS is particularly important in black-box optimization, where the a priori information about the problem instance is typically very limited, and where algorithms accumulate important insight during the optimization process itself. In our project, we will develop model-free and model-based AOS strategies for numerical black-box optimization. We will develop the techniques within a modular framework of the Covariance Matrix Adaptation Evolution Strategy (CMA-ES), and we will cross-validate our approaches on other state-of-the-art modular algorithm frameworks.
This PhD research project has been submitted for a funding request to “Sorbonne Center for Artificial Intelligence (SCAI)”. The PhD candidate selected by the project leader will therefore participate in the project selection process (including a file and an interview) to obtain funding.