Help with Neural Network - Transfer function

Hi all,

Just a question, while designing a neural network, how does one go about deciding on what particular transfer function to use for the network. I did go through some literature, but I could not find a document, that explains this. I mean, having a look at the transfer functions supported by matlab, how does one know, that "tansig" would be appropriate or "purelin" would be appropiate.

Is it just the matter of playing around with diff transfer functions and see which one suits best for the network, or is there actually some theory that explains which one can be used. A possible answer that came to my mind, was depends on the type of design you doing, but in respect to that, I came across this question.

Can someone provide some feedbacks and expertise comments

Regards

Perhaps i'm wrong, nn doesn't deal with transfer function. But the characteristic of the function which can be learn by nn. NN doesn't like math eq like transfer function. It's only learn by data in-out. These data maybe can be got from the t.f.

So we imitate the t.f's characteristic by the nn process learning.

I think you mean the function of each node in the neural network. In this case the choose of such function will effect your learning algorithm. in fact any nonlinear function will be as good as other if the learning algorithm is adjusted for such function. also there is no need for all nodes to have the same function.

there is a good bbok by raol rojas you can find one chapter that talks about this issues on his web site. But sorry i did not remember the site. Just search for the name raol rojas and neural networks.

i think he is referring to the activation function of the neural network nodes,
the best book for you to read is crc press handbook of neural networks,

In general, a neural network can be defined as weighted directed graphs in which neurons are nodes and directed weighted edges are the connection between neuron output and neuron inputs. The feedforward neural network (FFNN) is a neural network in which graphs have no loops . Each neuron performs some nonlinear multi input single output function (activation function or transfer function). The learning problem is to find the optimal combination of weights so that the network function  approximate a given function f, which is given implicitly through some examples (training set), as closely as possible.