Dear Jeff,
I tried the following example in the Encog Workbench, using the Analyst Wizard:
I used the described sunspot example, just instead of FeedForward network, I used the PNN/GRNN network. I kept all other parameters as described in the sunspot example. When starting the training, it never gets finished. When I checked the EGA file, I saw this architecture:
architecture=?->R(kernel=gaussian)->0
So, the number of output neurons is 0, because in the code in "AnalystWizard.java", method "generatePNN", it sets the number of output neurons to the number of classes, which doesn't seem to work for regression. When I replaced this by hard coding this to the number of predictable values (Lead Count), it seems to finish the training, although, I'm not sure, if this is the correct solution.
Dear Jeff, I tried the following example in the Encog Workbench, using the Analyst Wizard: I used the described sunspot example, just instead of FeedForward network, I used the PNN/GRNN network. I kept all other parameters as described in the sunspot example. When starting the training, it never gets finished. When I checked the EGA file, I saw this architecture: architecture=?->R(kernel=gaussian)->0 So, the number of output neurons is 0, because in the code in "AnalystWizard.java", method "generatePNN", it sets the number of output neurons to the number of classes, which doesn't seem to work for regression. When I replaced this by hard coding this to the number of predictable values (Lead Count), it seems to finish the training, although, I'm not sure, if this is the correct solution.
best regards, Winfried