2

I would appreciate a some insights into the workings of the PyBrain's neural network. I have a dataset of different household features that correspond to a certain household income. The task is to create a regression based on neural networks to be able to predict the income for given features.

I've tried the simple constructor 

pybrain.tools.shortcuts.buildNetwork(feature_count, 12, 1, recurrent=False)

and it kinda works. But if i change the hiddenlayer to use GaussianLayer or LinearLayer i am getting the NaNs as output during the training phase.

Is there maybe something else that needs to be taken care of when using these layers (I am guessing maybe feature selection, when they correlate)?

Thanks

d56
  • 825
  • 1
  • 9
  • 26
  • If this is something you still care about you need to include a sample of the data and more code showing what's failing. – Firestrand Mar 04 '14 at 02:57

1 Answers1

1

I solved a neural network regression problem using pybrain where I had to forecast the load on a power station using weather features. This appears to be the same problem as yours, except in application. I followed the guide here: http://fastml.com/pybrain-a-simple-neural-networks-library-in-python/ which brought me 90% of the way towards the final solution. I had 8 inputs and one outputs.

One "gotcha" I found was that I had to normalise my input values to 0 -> 1. The MSE value would not decrease on each EPOCH otherwise. Also, if any of my input vaues were NaN, I got continuous Nan values out.

I hope this helps.

Diarmaid O Cualain
  • 300
  • 2
  • 11