I am trying to do Statistical Downscaling of GCM outputs using ANN in R program. Initially i have standardised the predictors and find Principal components of predictors by Principal component analysis. Now i am finding out relation between principal components of Predictors and Observed data (predictand) by ANN.
When i have searched R code of ANN online , i have found that in the code instead of applying neuralnet function directly to the data, they are scaling data as follows:
scaled
Please check the link might be helpful for you.
http://www.iaeng.org/publication/IMECS2011/IMECS2011_pp356-361.pdf
Yes, you should scale your predictors again, even after PCA. Each principal component will have a different variance, and, to train a NN, you expect your predictors to have scaled unit variance. The need of scaling Y depends on the activation function you use on the last layer. If it is a linear activation, than no. If it is a sigmoidal activation, your predictand should be bound between -1 and 1.
- Badges
- Science topic
- Similar topics
- Linguistics
- Composition Studies
- Publication