Neural network for fitting histograms - tips for optimisation / setup?

Hi, I am currently using a FITNET to estimate underlying parameters for histogram data. As I am not greatly experienced in using neural networks, I stumbled across some problems (mainly that the network is not very good at predicting the parameters).

.

DATA+BACKGROUND

I want that the neural network takes a histogram (always between x-value 0 and 1 and normalized to 1) and gives me the underlying parameter that created this shape. I have the advantage that I can simulate these histograms for a multitude of parameter combinations (although that takes a long time - that is why I want to use the network in the first place and not just FMINSEARCH to compare the histograms between my simulations and the experiment).

.

NETWORK

Input size: 41 (=curve length, in my case 41 bins of a histogram), Hidden layer: 1 of size 10, Output parameter: 6, Training data size: around 500k. Otherwise the standard fitnet settings. See image below.

.

RESULTS

• Very bad regression (see image at the end). The outputs are in the right range, or display the correct trend - but are in general very inaccurate.
• I tried more/less hidden nodes - more increases the performance slightly
• trainbr slightly better than other training methods
• I tried the network with a simple fitting problem and it worked without a problem (in my case I just simulated multiple double gaussians with different positions and widths and the network would give me the peak positions and sigmas)

So - do you spot any general error. Or might my training set just be faulty? I could imagine some output parameter might confuse the network since their effect on the shape of the histogram is more complex. Unfortunately just reducing the number of parameter is not really an option. Thank you very much for your help (even just stating that if I set up the fitnet in a correct way would help me as then I would need to take a closer look at my training data I guess).