R 中的神经网络 package 大错误

Question

I am trying to figure out how to make the neuralnet package to work. I did some tests with data I created and with their outcomes (about 50 rows of data and three columns with the fourth being the outcome I wanted and it was made from simple mathematical executions like summing the other three columns) and so far so good. Then I decided to apply the package on real data. I downloaded the mpg dataset from here http://vincentarelbundock.github.io/Rdatasets/datasets.html

I was running the code bellow:

net<- neuralnet(cty~displ+year+cyl+hwy,
                datain, hidden=3)

Even if I have 3 hidden layers, or 8 or 18 the error is the same and the time that the package processes the data is relatively small from this amount of data (234 lines):

        Error Reached Threshold Steps
1 2110.173077    0.006277805853    54

Any good advice for this?

Answer 1

It's a scale problem i guess, you can normalize or scale it. There are differences between scaling and normalizing , it will affect your results and worths a separate question on SO:

normalize inputs

norm.fun = function(x){ 
  (x - min(x))/(max(x) - min(x)) 
}

require(ggplot2) # load mpg dataset
require(neuralnet)

data = mpg[, c('cty', 'displ', 'year', 'cyl', 'hwy')]
data.norm = apply(data, 2, norm.fun)

net = neuralnet(cty ~ displ + year + cyl + hwy, data.norm, hidden = 2)

Then you can denormalize the data

# restore data 
y.net = min(data[, 'cty']) + net$net.result[[1]] * range(data[, 'cty'])
plot(data[, 'cty'], col = 'red')
points(y.net)

scale inputs

data.scaled = scale(data)
net = neuralnet(cty ~ displ + year + cyl + hwy, data.scaled, hidden = 2)

# restore data 
y.sd = sd(data[, 'cty'])
y.mean = mean(data[, 'cty'])

y.net = net$net.result[[1]] * y.sd + y.mean
plot(data[, 'cty'], col = 'red')
points(y.net)

You can also try the nnet package, it's very fast:

require(nnet)

data2 = mpg
data2$year = scale(data2$year)
fit = nnet(cty ~ displ + year + cyl + hwy, size = 10, data = data2, linout = TRUE)
plot(mpg$cty)
points(fit$fitted.values, col = 'red')

Answer 2

Have you tried enabling linear output in your neural net model through linear.output=TRUE label?

neuralnet(formula, data, hidden = 1, threshold = 0.01,
  stepmax = 1e+05, rep = 1, startweights = NULL,
  learningrate.limit = NULL, learningrate.factor = list(minus = 0.5,
  plus = 1.2), learningrate = NULL, lifesign = "none",
  lifesign.step = 1000, algorithm = "rprop+", err.fct = "sse",
  act.fct = "logistic", ****linear.output = TRUE****, exclude = NULL,
  constant.weights = NULL, likelihood = FALSE)