How to model a neural network through the use of caret, R

Question

So I have a dataset that I've been performing machine learning algorithms on. I've performed MLR, stepwise regression, SVM and Random Forest on a dataset that is 180 x 160. I'm modelling one variable against 159 other variables, with 179 cases. It's all regression modelling. I've been using the caret package in which I use the train function to do 10 fold cross validation 10 times with the different machine learning algorithms. I was told to read up a paper that had used neural network models instead and got better results, so I've been trying to find a way of doing the same thing but with a neural network model instead.

I've had a look at doing the following:-

model <- train(RT..seconds.~., data = cadets, method = "AMORE", trControl = ctrl)

but it doesn't work. I was told that it wouldn't work as the train function does not have the AMORE packaged wrapped yet. So I looked to use nnet instead:-

model <- train(RT..seconds.~., data = cadets, method = "nnet", trControl = ctrl)

which worked. However, the RMSE value I got was 171, and when I looked at my predicted vs observed values, the predicted values were all just 1s and 0.9999s. Does anyone know what I'm doing wrong?

thanks!

Solution

You need to use the option linout = TRUE for the nnet function:

model <- train(RT..seconds.~., data = cadets, 
               method = "nnet", trControl = ctrl,
               linout = TRUE)

If you do not, a sigmoidal activation function is used and all of the predictions will be constrained to be on [0, 1].

Comments

[user2062207]

Thank you, it seems to have work! Although the answers I get are like grouped (like it deviates between 8 different numbers) but hell it still works, thanks