Combining Gaussian Processes with Neural Networks for Active Learning in Optimization

0555604 - ÚI 2022 RIV DE eng C - Conference Paper (international conference)
Růžička, J. - Koza, J. - Tumpach, J. - Pitra, Z. - Holeňa, Martin
Combining Gaussian Processes with Neural Networks for Active Learning in Optimization.
IAL@ECML PKDD 2021: Workshop on Interactive Adaptive Learning. Proceedings. Aachen: Technical University & CreateSpace Independent Publishing, 2021 - (Krempl, G.; Lemaire, V.; Kottke, D.; Holzinger, A.; Hammer, B.), s. 105-120. CEUR Workshop Proceedings, 3079. ISSN 1613-0073.
[IAL 2021: Workshop on Interactive Adaptive Learning /5./. Bilbao / virtual (ES), 13.09.2021-13.09.2021]
R&D Projects: GA ČR(CZ) GA18-18080S
Grant - others:Ministerstvo školství, mládeže a tělovýchovy - GA MŠk(CZ) LM2018140
Institutional support: RVO:67985807
Keywords : active learning * black-box optimization * artificial neural networks * Gaussian processes * covariance functions
OECD category: Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)
http://ceur-ws.org/Vol-3079/ial2021_paper9.pdf

One area where active learning plays an important role is black-box optimization of objective functions with expensive evaluations. To deal with such evaluations, continuous black-box optimization has adopted an approach called surrogate modelling or metamodelling, which consists in replacing the true black-box objective in some of its evaluations with a suitable regression model, the selection of evaluations for replacement being an active learning task. This paper concerns surrogate modelling in the context of a surrogate-assisted variant of the continuous black-box optimizer Covariance Matrix Adaptation Evolution Strategy. It reports the experimental investigation of surrogate models combining artificial neural networks with Gaussian processes, for which it considers six different covariance functions. The experiments were performed on the set of 24 noiseless benchmark functions of the platform Comparing Continuous Optimizers COCO with 5 different dimensionalities. Their results revealed that the most suitable covariance function for this combined kind of surrogate models is the rational quadratic followed by the Matérn 25 and squared exponential. Moreover, the rational quadratic and squared exponential covariances were found interchangeable in the sense that for no function, no group of functions, no dimension and combination of them, the performance of the respective surrogate models was significantly different.
Permanent Link: http://hdl.handle.net/11104/0330068