Optimized Temporal Multiplexing for Reservoir Computing with a Single Delay-Coupled Node
Hazem Toutounji, Johannes Schumacher, Gordon Pipa · IEICE Proceeding Series · 2014
The computational performance of reservoir computers based on a single delay-coupled node is critically dependent on the temporal multiplexing of input to the reservoir. Here we present an optimization of the temporal multiplexing by means of optimizing virtual node distance to maximize the response of the delay-coupled system to stimulation. After demonstrating the analytical approach, we discuss how the optimization has a single optimum (concave problem), and illustrate the improvement of the reservoir computer's performance. To this end we predict a NARMA-10 time series and show that optimizing temporal multiplexing reduces the normalized root mean squared error by ∼ 8%.