基于储备池计算的大采样间隔混沌系统的学习与预测

Reservoir computing is an efficient artificial neural network with low training cost and low hardware overhead. It is widely used in time sequence information processing, such as waveform classification, speech recognition, time series prediction, etc. However, in practical applications, researchers can only use limited information from the system for predictions, and the sampling interval cannot be adjusted freely due to the limitations of the actual system. Based on the above situation, we demonstrate the impact of time and space sampling intervals on the short-term and long-term prediction capabilities of the reservoir computing and compare it with the existing numerical methods. It can be found that for chaotic systems, the reservoir computing can learn and reproduce the systems' states at almost five times larger spatio-temporal intervals compared to classical numerical methods, such as fourth-order Runge-Kutta and spectral methods. Our results show the captivity of reservoir computing in the applications with limitation of spatio-temporal intervals, and pave the way to reservoir-based fast numerical simulation methods.