Generalized regression neural network (GRNN) can be used for regression, prediction, and classification, so: what are the drawbacks of the GRNN?, and how can we overcome these drawbaks?!
Its size can be huge, which would make it computationally expensive.
There is no optimal method to improve it.
There are several ways to overcome this disadvantage. One is to use a
clustering versions of GRNN or precisely, a double clustering version because one use of the clustering algorithm is not sufficient. The second clustering not only speeds up the testing but also replaces the division required for kernel regression with simply the search for the nearest neighbor.
GRNN are excellent predictors for a small dataset with low correlation. The notable disadvantages of this network are:
1. Optimal value selection for the 'smoothing parameter' (various analytical and evolutionary algorithms are available to do that).
2. Poor generalization ability for highly correlated data.
3. Scalability issue for high dimensionality data.
4. Sensitivity to noise in data and sudden varying attributes in data.
5. Required a hidden neuron for each data packet causing an increase in training time for huge datasets.
Recommendations:
Use optimization algorithm for the setting of smoothing parameter, use normalization & some sort of regularization, reduce data dimensionality by ICA/PCA, use novelty packet information for dataset size reduction.