I am trying to fit a neural net in R, the goal is to predict an integer value. I have a lot of covariates, some are continuous, some are categorial and I standardized these. When I use the nnet function from the nnet package with one hidden layer and 5 nodes there, it works kind of okay. But I don't like that I cannot see what the learning function, the act. function, ... is. When using mlp from the RSNNS package with the same number of hidden neurons, it gives me constant outputs (which is wrong) and the function is very slow. I thought the functions should give quite similar results as the constructed network should be the same.
nn <- nnet(x_train, y_train,size=5, linout=T, maxit = 50, MaxNWts=16350)
nn_p <- predict(nn, newdata=x_test)
> nn_p
y
0.4484153002
-1.2744682375
0.4484153002
0.4657251465
0.4484153002
0.4484153002
0.4484153002
0.4484153002
-1.2744682375
nn3 <- mlp(x_train, y_train, size=c(5), maxit=50, linOut = TRUE,
hiddenActFunc = "Act_TanH", learnFunc = "BackpropMomentum")
nn3_p <- predict(nn3, x_test)
> nn3_p
y
-0.2601733208
-0.2601733208
-0.2601733208
-0.2601733208
-0.2601733208
-0.2601733208
-0.2601733208
-0.2601733208
-0.2601733208