The online optimization of brain-computer interface stimulus parameters, a simulation

Keywords
Loading...
Thumbnail Image
Issue Date
2023-08-20
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
A brain-computer interface operates by presenting stimuli to elicit brain responses that can be acted upon. As the discriminability of these responses is subject to the stimulus characteristics, the presentation of the optimal stimuli could improve the performance of the interface. Unfortunately, these optima are subject-specific. Furthermore, the optimization of said stimuli is complicated by the potentially high level of non-independent and identically distributed noise that comes with the recorded responses. The aim of this thesis is to demonstrate that these challenges can be alleviated by introducing the optimization process with heteroskedatic modelling and the replication of existing designs. To this end, various Bayesian optimization algorithms are tested on simulations that are designed to resemble the aforementioned online optimization objective. The compared optimization algorithms differ in the surrogate models, replication schemes and selection strategies that have been used. As it is impossible to change the stimulus characteristics that have been used to record existing datasets, the simulated objective functions model the response discriminability as a function of the decoding parameters. To assess the effectiveness of the optimization algorithms, the algorithms are tested on one, two and seven-dimensional objective functions. The results suggest that a Bayesian optimization algorithm that is enriched with heteroskedastic Gaussian process regression and the locationbased evaluation of existing designs significantly outperforms the random sampling baseline for all objective functions that have been considered. Even though the thesis is focused at braincomputer interfaces, application of the results is widespread since heteroskedastic noise can also be encountered with black-box optimization within other domains.
Description
Citation
Faculty
Faculteit der Sociale Wetenschappen