Doubly Trained Evolution Control for the Surrogate CMA-ES

0466878 - ÚI 2017 RIV CH eng C - Conference Paper (international conference)
Pitra, Zbyněk - Bajer, L. - Holeňa, Martin
Doubly Trained Evolution Control for the Surrogate CMA-ES.
Parallel Problem Solving from Nature - PPSN XIV. Cham: Springer, 2016 - (Handl, J.; Hart, E.; Lewis, P.; López-Ibáñez, M.; Ochoa, G.; Paechter, B.), s. 59-68. Lecture Notes in Computer Science, 9921. ISBN 978-3-319-45822-9. ISSN 0302-9743.
[PPSN XIV. International Conference on Parallel Problem Solving from Nature /14./. Edinburgh (GB), 17.09.2016-21.09.2016]
R&D Projects: GA MZd(CZ) NV15-33250A
Grant - others:ČVUT(CZ) SGS14/205/OHK4/3T/14; GA MŠk(CZ) ED2.1.00/03.0078; GA MŠk(CZ) LO1611; GA MŠk(CZ) LM2010005
Institutional support: RVO:67985807
Keywords : black-box optimization * surrogate model * evolution control * Gaussian process
Subject RIV: IN - Informatics, Computer Science

This paper presents a new variant of surrogate-model utilization in expensive continuous evolutionary black-box optimization. This algorithm is based on the surrogate version of the CMA-ES, the Surrogate Covariance Matrix Adaptation Evolution Strategy (S-CMA-ES). Similarly to the original S-CMA-ES, expensive function evaluations are saved through a surrogate model. However, the model is retrained after the points in which its prediction was most uncertain have been evaluated by the true fitness in each generation. We demonstrate that within small budget of evaluations, the new variant of S-CMA-ES improves the original algorithm and outperforms two state-of-the-art surrogate optimizers, except a few evaluations at the beginning of the optimization process.
Permanent Link: http://hdl.handle.net/11104/0265826