问题
I am trying to implement a customized error function in package neuralnet in R.
Normally ’sse’ and ’ce’ which stand for the sum of squared errors and the cross-entropy are used to calculate error.Can anyone provide me details about how to implement own error function. Though the package says we can use customized error function,there is no help in the user Manuel about this.
回答1:
I had the same Problem. This is the solution/help I received. You can use the usual definition of R functions (function(x,y){...}). Hence, the error function must be of the type function(x,y) where x is the fitted value and y is the true value.
Please refer to the following example.
library(neuralnet)
AND <- c(rep(0,7),1)
OR <- c(0,rep(1,7))
binary.data <- data.frame(expand.grid(c(0,1), c(0,1), c(0,1)), AND, OR)
set.seed(3)
print(net <- neuralnet(AND+OR~Var1+Var2+Var3, binary.data, hidden=0, rep=10, err.fct="sse", linear.output=FALSE))
#Call: neuralnet(formula = AND + OR ~ Var1 + Var2 + Var3, data = binary.data, hidden = 0, rep = 10, err.fct = "sse", linear.output = FALSE)
#
#10 repetitions were calculated.
#
#Error Reached Threshold Steps
#7 0.04043122185 0.008248439644 116
#5 0.04426319054 0.009619409680 124
#8 0.04698485282 0.007947430014 117
#2 0.04931335384 0.008792873261 88
#1 0.04965332555 0.009631079320 89
#4 0.05396400022 0.009092193542 96
#6 0.05488395412 0.009990028287 124
#3 0.06383087672 0.009964206587 94
#10 0.51657348285 0.008602371325 51
#9 0.52514202592 0.007890927099 40
set.seed(3)
custom <- function(x,y){1/2*(y-x)^2}
print(net <- neuralnet(AND+OR~Var1+Var2+Var3, binary.data, hidden=0, rep=10, linear.output=FALSE, err.fct=custom))
#Call: neuralnet(formula = AND + OR ~ Var1 + Var2 + Var3, data = binary.data, hidden = 0, rep = 10, err.fct = custom, linear.output = FALSE)
#
#10 repetitions were calculated.
#
#Error Reached Threshold Steps
#7 0.04043122185 0.008248439644 116
#5 0.04426319054 0.009619409680 124
#8 0.04698485282 0.007947430014 117
#2 0.04931335384 0.008792873261 88
#1 0.04965332555 0.009631079320 89
#4 0.05396400022 0.009092193542 96
#6 0.05488395412 0.009990028287 124
#3 0.06383087672 0.009964206587 94
#10 0.51657348285 0.008602371325 51
#9 0.52514202592 0.007890927099 40
You can use basically every error function that can be differentiated.
来源：https://stackoverflow.com/questions/25510960/how-to-implement-own-error-function-while-using-neuralnet-package-in-r