Generating prediction using a back-propagation neural network model on R returns same values for all observation

问题

I'm trying to generate prediction using a trained backpropagation neural network using the neuralnet package on a new data set. I used the 'compute' function but end up with the same value for all observations. What did I do wrong?

# the data
Var1 <- runif(50, 0, 100)
sqrt.data <- data.frame(Var1, Sqrt=sqrt(Var1))

# training the model
backnet = neuralnet(Sqrt~Var1, sqrt.data, hidden=2, err.fct="sse", linear.output=FALSE, algorithm="backprop", learningrate=0.01)

print (backnet)

Call: neuralnet(formula = Sqrt ~ Var1, data = sqrt.data, hidden = 2,     learningrate = 0.01, algorithm = "backprop", err.fct = "sse",     linear.output = FALSE)

1 repetition was calculated.

        Error Reached Threshold Steps
1 883.0038185    0.009998448226  5001

valnet = compute(backnet, (1:10)^2)

summary (valnet$net.result)

      V1           
Min.   :0.9998572  
1st Qu.:0.9999620  
Median :0.9999626  
Mean   :0.9999505  
3rd Qu.:0.9999626  
Max.   :0.9999626  

print (valnet$net.result)

         [,1]
[1,] 0.9998572272
[2,] 0.9999477241
[3,] 0.9999617930
[4,] 0.9999625684
[5,] 0.9999625831
[6,] 0.9999625831
[7,] 0.9999625831
[8,] 0.9999625831
[9,] 0.9999625831
[10,] 0.9999625831

回答1:

I was able to get the following to work:

library(neuralnet)

# the data
Var1 <- runif(50, 0, 100)
sqrt.data <- data.frame(Var1, Sqrt=sqrt(Var1))

# training the model
backnet = neuralnet(Sqrt~Var1, sqrt.data, hidden=10, learningrate=0.01)

print (backnet)


Var2<-c(1:10)^2

valnet = compute(backnet, Var2)

print (valnet$net.result)

Returns:

     [,1]
[1,] 0.9341689395
[2,] 1.9992711472
[3,] 3.0012823496
[4,] 3.9968226732
[5,] 5.0038316976
[6,] 5.9992936957
[7,] 6.9991576925
[8,] 7.9996871591
[9,] 9.0000849977
[10,] 9.9891334545

According to the neuralnet reference manual, the default training algo for the package is backpropogation:

neuralnet is used to train neural networks using backpropagation, resilient backpropagation (RPROP) with (Riedmiller, 1994) or without weight backtracking (Riedmiller and Braun, 1993) or the modified globally convergent version (GRPROP) by Anastasiadis et al. (2005). The function allows flexible settings through custom-choice of error and activation function. Furthermore the calculation of generalized weights (Intrator O. and Intrator N., 1993) is implemented.

来源：https://stackoverflow.com/questions/19206052/generating-prediction-using-a-back-propagation-neural-network-model-on-r-returns

标签

neural-network

backpropagation