I'm using caret package and 'neuralnet' model so as to find the best tuning parameters for a neural network based on a data set which contains several predictors transformed by PCA. This data set also contains two output numeric variables, so I want to model these two variables against the predictors. Thus, I'm performing regression.
When using the package 'neuralnet', I get the desired output: a network whose output layer consists of two neurons, corresponding to the two output variables that I want to model, as you can see from the following code.
library(neuralnet)
neuralnet.network <- neuralnet(x + y ~ PC1 + PC2, train.pca.groundTruth, hidden=2, rep=5, algorithm = "rprop+", linear.output=T)
> head(compute(neuralnet.network, test.pca[,c(1,2)])$net.result)
[,1] [,2]
187 0.5890781796 0.3481661367
72 0.7182396668 0.4330461404
107 0.5854193907 0.3446555435
228 0.6114171607 0.3648684296
262 0.6727465772 0.4035759540
135 0.5559830113 0.3288717153
However, when using the same model with train function from caret package, the output consists of just one single variable, named '.outcome', which is in fact the sum of the two variables. This is the code:
paramGrid <- expand.grid(.layer1 = c(2), .layer2 = 0, .layer3 = 0)
ctrl <- trainControl(method = "repeatedcv", repeats = 5)
set.seed(23)
caret.neuralnet <- train(x + y ~ PC1 + PC2, data = train.pca.groundTruth, method = "neuralnet", metric = "RMSE", tuneGrid = paramGrid, trControl = ctrl, algorithm = "rprop+", linear.output = T)
> head(predict(caret.neuralnet, test.pca[,c(1,2)]))
[1] 0.9221328635 1.1953289038 1.0333353272 0.9561434406 1.0409961115 0.8834807926
Is there any possibility to prevent caret train function from interpreting the symbol '+' in a formula as summation but as the specification of several output variables, just as neuralnet does? I've tried the x-y form, though it doesn't work.
I would like to know whether there is any form to do that without training separate models for each output variable.
Thank you so much!
