Closed PetrToman closed 12 years ago
Jeff checked in code that eliminates this error and allows PNN to be used with the workbench. However, it currently fails to converge with the Iris dataset, so we are looking at that.
I fixed a bug that was causing a total failure to converge. Now PNN will train Iris to a flat zero error. And this is actually correct. However, it is massively "overfit", as the evaluation set performs at about a 23% error. I am not sure if this is just the nature of PNN, since it does store the training data. I am assigning to Jeff to look and see if the algorithm is functioning as it should, and the network is not just memorizing.
The original issue reported has been fixed. But I would still like to take a look at overfitting with PNN. Created Issue #87 for that, for Encog 3.2.
Run Analyst on data1.csv (http://dione.zcu.cz/~toman40/encog/data1.zip) and choose "PNN/GRNN Network". The following exception will be thrown: