Hello, I m trying to develop a Feed Forward Network with Backpropogation. Initially i developed it for training simple cases like all possible boolean functions of 2 inputs. For that i developed a feedforward network with 2-inputs, 1 hidden layer with 2 neurons and an output layer with single neuron. I used backpropogation learning algorithm. The problem i m having is that i dont know what value of the 'learning rate' to use. I tried many values form 0.1 to 1.0 but still my network is unable to correctly classify some of the very simple linearly separable cases. Out of 16 possible cases i was able to correctly classify about 9 input patterns in the training set( all 9 were linearly separable cases ). Is this normal to get such results on such simple training data. Or may be i m doing some mistake in my program. I havent used techniques like momentum and weight elimination yet but my question is that,is it possible to not get 100% correct results on training data as simple as this when not using momentum or weight elimination. Actually i want to test my network on boolean functions before training it from other more complex data set. Any tip will be appreciated. Thanks Tariq

What is your activation function? (Is it one of the typical sigmoids?)

Did you test training rates smaller than 0.1?

Yes im using sigmoid function at every neuron as my activation function,

1/(1+e^(-x))

I also tried using many values of training rates including values smaller than 0.1 but still its not working. Initially i assigned random weights and biases to all connections b/w 0 and 1. does initial values of weights and biases have any effect on the way traning proceeds?

Thanks

never mind i found a mistake in my program. I was using wrong weights for calculating Delta. Its corrected and now my network is performing 100% correctly on boolean functions of 2-input.

Thanks for ur reply
Bye