How can i train a neural network already optimized with GA in MATHLAB with train net

I already optmized a neural network using GA in MATHLAB, but i believe this has only achieved an optimum region, so i want to further train the same network using the train net in order to extend the search. or is the neural network already a GANN, so no need for that. pls i need to know. Thanks

Answer 1

[Greg Heath]

1. MATLAB is named for MATRIX algorithms: THERE IS NO "H"

2. If you need help why are you giving up as little information as possible? Are you working on something you plan to patent? I'm warning you now: If you patent anything with my ideas not compensated for, you are going to be in BIG, BIG trouble.

3. Now, down to busine$$:

""I already optmized a neural network using GA in MATHLAB,..."

What kind of network? Why that one? What size input and target matrices? What topology? What hidden and output layer transfer functions? What performance function? If regression or curve-fitting, how much of the target variance were you able to model for the training, validation and testing subsets? If classification or regression, what error rates for each class in the trn/val/tst subsets.

4. It sounds like you are afraid that you have reached a local optimum of the performance function that is significantly different from the global optimum .... IS THAT CORRECT?

5. My philosophy is

 a. Specify a performance goal (e.g., model 99.9% of the average target variance) 
 b. Systematically vary input parameters (e.g, number of hidden nodes and initial weights) 
 c. Design multiple models (e.g., 10 to 100, depending on the problem).
 d. Obtain summary performance statistics of the trn/val/tst subsets.
 e. Rank the nets via val set performance stats.
 f. Obtain UNBIASED performance estimates on unseen data using the tst set performance stats.

In short, I think spending time on multiple designs is much more fruitful than trying to beat one design into submission.

Hop this helps.

Greg

[Greg Heath]

It would be interesting to time GA vs MRILM (Multiple Random Initialization Levenberg-Marquardt).