An Optimized Physical Reservoir Computing Through Genetic Algorithm

Reservoir Computing (RC), a compact recurrent neural network (RNN), is an efficient artificial neural network suitable for processing timing signals. In our previous work on a physical reservoir consisting of single-walled carbon nanotubes (SWCNTs) network, its computing performance was presented to be improved by phosphomolybdic acid-modification on CNTs. In this paper, a genetic algorithm was used to search some hyperparameters related to its computing architecture, including input gain, input position, auxiliary input, regularization coefficient and leak rate, for two test benches, i.e., NARMA10 time series prediction and epileptic seizure EEG classification. The parametric setting for improved computing capability of this physical system were found. The best performances were achieved at normalized root-mean-square error of 0.0877 for NARMA10 and an accuracy of 98.33% for EEG classification, respectively.