Although this publication is mine, I think GRNN is not the appropriate tool to make a classification task. Results are good anyway, but we tried to make a functional mapping from features to the values 1, 2 and 3. A more correct and powerful approach is to use PNN (Probabilistic Neural Networks). This kind of network is intended for classification tasks.
Conference Paper SEGMENTACIÓN DE IMÁGENES CEREBRALES DE RESONANCIA MAGNÉTICA ...
You're right. However PNN have their own significant drawbacks. In my opinion, to solve your problem, it is desirable to use associative artificial intelligence systems (machines). They will take into account the diversity of possible factors and the relationship between them.
I do not think that there is an optimal classifier (in terms of accuracy), as they are all approximate methods, and they are using sample data, which is a very small proportion of the real population, therefore, in the absence of some critical data, the "best" classifiers become useless.
I bit on GRNN to be better than PNN if we change the training data, and vice versa.
this is supported by the no-free-lunch theorem and my dear K.S. Madhavan conclusion above.
In my experience, GRNN networks work quite well when the data are properly preprocessed.
- Badges
- Science topic