[comp.ai.neural-nets] NN solution of non-deterministic problems.

sra@idx.uucp (08/07/90)

Let me add $0.02 about non-deterministic learning on NN.  My quick half-thought
on the learning about rainfall is "it depends".

If the learning takes place with usual back propagation using sigmoid function,
the error propagation is based on least-square-error.  That would suggest that
the output might resemble an RMS kind of solution.

It further depends on the "random" order in which the learning is done and
whether the modification of weights is done after a pass or after each case. 
You might get stuck with the "last wins" problem.

	steve