Master theses
Gaussian Processes and Neural Networks as Surrogate Models for the CMA Evolution Strategy
Author
Jiří Růžička
Year
2021
Type
Master thesis
Supervisor
Ing. Jan Koza
Reviewers
prof. Ing. RNDr. Martin Holeňa, CSc.
Department
Summary
This thesis is about surrogate models for Doubly Trained Covariance Matrix Adaptation Evolution Strategy, which is a modification of CMA-ES in black-box optimisation. We use Gaussian Processes and Neural Networks to model the objective function and to decrease the number of objective function evaluations. Especially, we wanted to explore whether the combination of Gaussian Processes and Neural Networks in the form of Multilayered Perceptron will outperform Gaussian Processes alone. We created experiments on COCO-BBOB testbed to compare the performances of particular surrogate models. The statistical significance of all measured results is thoroughly verified using the Friedman test, followed by multiple comparison post-hoc tests. The thesis did not find that the combination would perform better and found the Gaussian Processes without Multilayered Perceptron as a better surrogate model for Doubly trained CMA-ES in black-box optimisation.