Re: Help on selection of optimal smoothing parameter for Probabilistic Neural Networks

In almost all cases like this, optimal parameter selection must be
done using cross validation.  This is a scheme in which you separate
your training data into multiple pieces so that you can train multiple
times for different values of your parameter evaluating performance
for each different parameter value on the held out training data.  You
should avoid using your test data for this parameter selection
process, of course, to get a realistic performance estimate.

In addition, it is also common to use multi-way cross validation to
get some idea about the variability of parameters selected this way. 
On example would be to divide your training data into 5 parts holding
each of the 5 parts out in turn from the other 4.  This would give you
5 estimates for the optimal parameter as well as 5 estimates of the
optimal performance.  With this information, you can select a
consensus estimate for the parameter, train on all of your training
data and then evaluate on your test data.

My own recommendation is generally to avoid non-linear systems like
neural nets until you have exhausted the potential of simpler
techniques such as logistic regression.  IF you must have
non-linearity, it can often be introduced in the form of a simple
composite variable in a logistic regression allowing you to retain the
convergence guarantees of the simpler regression while matching the
performance of a neural net.

For that matter, feature detection and processing is key.  You can
ruin any good regression system by giving it perverse features.

[EMAIL PROTECTED] (Dr. V. Ravi) wrote in message news:<[EMAIL PROTECTED]>...
> Hi all:
> 
> Could anyone let me know the source/paper/link where methods for
> selection of the optimal smoothing parameter for probabilisitc neural
> networks (PNN) are suggested?
> 
> It is an urgent requirement.
> 
> Thank in advance,
> 
> Dr. Ravi