Task-Independent Evaluation Metrics of Carbon Nanotube Based In-Materio Reservoir Computing

Reservoir computing (RC) is a very lightweight machine learning framework, suitable for edge information processing in the era of IOT. Recently, increasing efforts on RC to be implemented using various novel materials and devices are emerging, such as carbon nanotube, memristor, etc. However, how to configure the reservoir’s behavior suitable for a certain application is still an open problem. In this paper, our attention is paid to the behavior space of the reservoir including its all the currently available task-independent evaluation metrics, i.e., rank of kernel quality, rank of generalization capacity, information processing capacity, and memory capacity. A genetic algorithmis used to explore the space in order to identify the performance boundary of the reservoir. It is applied to characterize the computing capabilities of a carbon nanotube based in-materio reservoir system, especially to find its critical metric to implement the prediction of nonlinear series such as Nonlinear Autoregressive Moving Average with 10th order timelag (NARMA 10).