The general regression scenario is you want a deterministic multivariate function that yields an O-dimensional output vector given an I-dimensional input vector. Although the function is deterministic, it works well for nondeterministic inputs and outputs with moderate to high signal-to-noise ratios.
Typically, the O-dimensional output function is approximated by a weighted sum of H tanh functions whose arguments are linear combinations of the I-dimensional input vector components. The node topology of the net is I-H-O.
The weights are approximated, iteratively, with N pairs of input/output-target examples
[ I N ] = size(input)
[ O N ] = size(target)
The number of scalar equations obtained from the examples are
Neq = N*O
The number of unknown weights is
Nw = (I+1)*H+(H+1)*O
The 1s indicate bias weights connected to constant unit source.
To mitigate noise, interference and measurement errors, it is desired that
Neq >> Nw
and the approximation is obtained by minimizing the sum of the squared differences between the approximation output and the design target.
The double inequality is equivalent to H << Hub (upper bound) where
Hub = -1 + ceil( (N*O-O) / (I+O+1))
There are a myriad of details and exceptions that you can worry about after you have tried the examples in the documentation
help fitnet
doc fitnet
for regression and
help patternnet
doc patternnet
for classification.
To practice with more data try
help nndatasets
doc nndatasets
Hope this helps.
Greg
