Echo state networks in R
rESN is not available on CRAN, but can be installed using the R package devtools. rESN can be installed with the following R code:
devtools::install_github("jaredhuling/rESN")
library(rESN)U.train is the covariate matrix. Each row is an observation in time. Each column is a covariate. Y.train.noisy is a response in that depends nonlinearly on U over time. n.neurons defines the number of neurons in the "reservoir" (hidden layer). density is the density of non-zero connections in the reservoir. lambda is the ridge regression tuning parameter
data(esn_data)
net <- newESN(example_data$Y.train.noisy,
example_data$U.train,
n.neurons = 1000,
density = 0.02,
back.density = 0.02,
leak.rate = 0.6,
lambda = 10)
net <- train(net)
ypred <- predict(net, u = example_data$U.test)
range <- 431:480
all <- c(example_data$Y.test.noisy[range,1], example_data$Y.true.test[range,1],
ypred[range])
plot(example_data$Y.test.noisy[range,1], type = "l", lwd = 5, ylim = range(all), col = "green")
lines(ypred[range], col = "blue", lwd = 3)
lines(example_data$Y.test.signal[range,1], col = "black", lwd = 5)
legend("bottomleft", inset=0.01, legend=c("Noisy (observed) seq","ESN Prediction","True Signal"),
col=c("green", "blue", "black"), lwd = 3, cex = 0.75)