increasing nftool accuracy?

Hi all,

New to nftool. I have an application where I have two amplitudes in a fourier transform, and the ratio of the two produce a linear output from 0 to 1. I am using a two layer feed forward network with 4 hidden neurons and a single linear output. I have read that two peaks in an FT are best approximated by 4 hidden layes, one each for the rise and fall on each side of the Lorentzian peak. I used 2000 trainng sets with the output divided into 2000 steps from 01 to 1, with linewidths that changed randomly with a mean of 3 and standard devation of 1. My noise is 1/100th of the amplitude (random for each training set). This takes about 20 minutes to train on my computer. The training looks fantastic, with a .99 r-squared and very small noise in the answer (used default training and validation percentages in nftool). Looks like a major winner. then I test with 100 samples in a newly simulated data set and I get this.

The error is literally 20 to 40% higher than when I trained the data.

Any idea how to increase the accuracy of the network? It was fantastic in training, but a new set of noise values just blew it out of the water. I thought I had enough different noise samples (2000) to make a fairly robust network, but this is a fail.

Thanks