Reducing output matrix according to mean TE

The procedure for choosing output weights. First the training reservoir signals were collected and TE matrix was computed based on these signals. Next the mean TE of target unit was computed and only those units where chosen for output matrix computations which mean TE was higher then some Threshold. The threshold went from the minimum mean target TE to the maximum mean target TE in 100 steps. This was repeated 100 times. The network settings were same as in the IJCNN article for comparability reasons.

August 13, 2018August 13, 2018

Lyapunov exponent exploration revisited

This time with normalized (mean=0, variance=1) input + added a new input dataset (Multiple sinewave oscillator aka MSO)

July 18, 2018July 18, 2018

80: Sigma, Tau & Rho exploration

I have added spectral radius scaling of reservoir matrix to desired value $\rho$ along $\sigma$ and $\tau$ parameters in W $\sim$ N(0, $\sigma^2$ ) and $W^{in}$ $\sim$ Unif(- $\tau$ , $\tau$ ) to the exploration. $100$ neuron reservoir was used and $100$ instances for every pair ( $\tau$ , $\sigma$ ) were generated.

The Z values are mean NRMSE from 100 instances.

NARMA

NARMA without unscaled

Mackey – Glass

Mackey-Glass unscaled

Lorenz

Lorenz unscaled

Minimum achieved mean:
Rendered by QuickLaTeX.com

July 15, 2018July 15, 2018

79: W & Win exploration

Exploration of $\sigma$ and $\tau$ parameters in W $\sim$ N(0, $\sigma^2$ ) and $W^{in}$ $\sim$ Unif(- $\tau$ , $\tau$ ). $100$ neuron reservoir was used and $100$ instances for every pair ( $\tau$ , $\sigma$ ) were generated.