How to decide the range for weights and bias based on input data set ?

Based on input data, can we decide the range for weights and bias in neural network ? Like weight should be between (-1,1). Thank you

Answer 1

[Greg Heath]

You can but you do not have to. By default, variables are normalized to [-1,1] and the initial weights are automatically determined using the Widrow-Nguyen Algorithm. If you really want to, look up the logic for the algorithm.

If all weights are initially zero, train will automatically initialize them. If you train multiple nets in a loop, after the first pass the weights will be nonzero and train will just continue using the final weights obtained from the previous loop.

To prevent this, use the function configure to initialize weights before each call of train.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Greg Heath]

You might be interested in my thread "Nonsaturating Initial Weights" in comp.ai.neural-nets.

[Suresh]

Hi Greg, thank you very much for answer. Matlab neural networks uses Windrow-Nguyen algorithm for weight initialization. Still, if there is any formula to find range of weights and bias please share. Thank you