[R] nnet question

[Wensui Liu]

AM,
Don't worry. It is correct to get different correlation each time.
Unless you are very lucky, you will get different prediction for each
different training process, depending on your initial random state.

Take a look at Dr Ripley's MASS book, there are several excellent
examples on how to use nnet.

On 1/28/07, Aimin Yan <aiminy at iastate.edu> wrote:
> Hello,
> I use nnet to do prediction for a continuous variable.
> after that, I calculate correlation coefficient between predicted value and
> real observation.
>
> I run my code(see following) several time, but I get different correlation
> coefficient each time.
>
> Anyone know why?
>
> In addition, How to calculate prediction accuracy for prediction of
> continuous variable?
>
> Aimin
> thanks,
>
>
>  > m.nn.omega <- nnet(omega~aa_three+bas+bcu+aa_ss, data=training, size=2,
> linout=TRUE)
> # weights:  57
> initial  value 89153525.582093
> iter  10 value 15036439.951888
> iter  20 value 15010796.121891
> iter  30 value 15000761.804392
> iter  40 value 14955839.294531
> iter  50 value 14934746.564215
> iter  60 value 14933978.758615
> iter  70 value 14555668.381007
> iter  80 value 14553072.231507
> iter  90 value 14031071.223996
> iter 100 value 13709055.312482
> final  value 13709055.312482
> stopped after 100 iterations
>  > pr.nn.train<-predict(m.nn.omega,training)
>  > corr.pr.nn.train<-round(cor(pr.nn.train,training$omega),2)
>  > pr.nn.test<-predict(m.nn.omega,test)
>  > corr.pr.nn.test<-round(cor(pr.nn.test,test$omega),2)
>  > cat("correlation coefficient for train using neural
> network:",corr.pr.nn.train,"\n")
> correlation coefficient for train using neural network: 0.32
>  > cat("correlation coefficient for test using neural
> network:",corr.pr.nn.test,"\n")
> correlation coefficient for test using neural network: 0.39
>
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>


-- 
WenSui Liu
A lousy statistician who happens to know a little programming
(http://spaces.msn.com/statcompute/blog)