多層ニューラルネットワークのベイズ最適化による移動ロボット制御向けBCIの性能改善

抄録

The aim of this study is to realize practical classification performance of neural networks as an EEG-based BCI for mobile robot control by means of hyperparameter optimization in training the neural networks. The hyperparameters could be intuitively set, however, the classification performance will improve if you determine the hyperparameters in a more appropriate way. There are several methods for parameter optimization. Grid Search draws a grid on a parameter search space, examining all intersection representing a combination of the parameters. The method can find the best parameters with a high probability, but it takes a huge amount of time. In the authors' preceding study, Bayesian optimization was applied to training the neural networks in a shorter time and achieved better performance. In this study, the authors have more finely arranged the hyperparameter search space and executed Bayesian Optimization. As a result, the performance was further improved.