
174
Chapter 10
This chapter discusses the difficult problem of deciding on the
structure and training of a feedforward network. It is assumed that
the problem to be solved indicates that a multiple-layer feedforward
network is the most appropriate model. Given that assumption, we
must design the network. While there are no hard and fast rules for
defining network parameters, we can generally be safe if we follow
three guidelines:
1) Use one hidden layer.
2) Use very few hidden neurons.
3) Train until you can't stand it any more.
Let us now examine these guidelines more closely.
How Many Hidden Layers?
It has already been seen (page 85) tha