[R] problem with optim()

[Jonathan Daily]

What do you mean when you say "wrong results"? What do you expect for
the output? Your code doesn't work for me because it references X in
places and X is not defined.

Have you tested your functions to make sure they return reasonable values?

On Thu, May 19, 2011 at 9:17 AM, chirine wolley
<wolley.chirine at hotmail.com> wrote:
>
> Dear R-users,
>
> I would like to maximize the function g above which depends on 4 parameters (2 vectors, 1 real number, and 1 matrix)  using optim() and BFGS method. Here is my code:
>
> # fonction to maximize
>
> g=function(x)
> {
> x1 = x[1:ncol(X)]
> x2 = x[(ncol(X)+1)]
> x3 = matrix(x[(ncol(X)+2):(ncol(X)+1+ncol(X)*ncol(Y))],nrow=ncol(X),ncol=ncol(Y))
> x4 = x[(ncol(X)+1+ncol(X)*ncol(Y)+1):length(x)]
> res1=rep(0,nrow(X))
> res2=matrix(0,nrow=nrow(X),ncol=ncol(Y))
> log.res2=matrix(0,nrow=nrow(X),ncol=ncol(Y))
> res2.b=rep(0,nrow(X))
> res3 = rep(0,nrow(X))
> res3.b = rep(0,nrow(X))
> for (i in 1:nrow(X))
> {
> res1[i]=1/(1+exp(-t(x1)%*%X[i,]-x2))
> for (t in 1:ncol(Y))
> {
> res2[i,t] = ((1-(1+exp(-t(x3[,t])%*%X[i,]-x4[t]))^(-1))^(abs(Y[i,t]-Yb[i])))*(((1+exp(-t(x3[,t])%*%X[i,]-x4[t]))^(-1))^(1-abs(Y[i,t]-Yb[i])))
> log.res2[i,t]=log(res2[i,t])
> res2.b[i]=res2.b[i]+log.res2[i,t]
> }
> res3[i] = p_tilde[i]*log(res1[i])
> res3.b[i] = p_tilde[i]*(res2.b[i])
> }
> -(ncol(Y)*sum(res3)+sum(res3.b))
>
> }
>
> ##### Gradiants:
>
> gr=function(x)
> {
> x1 = x[1:ncol(X)]
> x2 = x[(ncol(X)+1)]
> x3 = matrix(x[(ncol(X)+2):(ncol(X)+1+ncol(X)*ncol(Y))],nrow=ncol(X),ncol=ncol(Y))
> x4 = x[(ncol(X)+1+ncol(X)*ncol(Y)+1):length(x)]
> gr1 = rep(0,ncol(X))
> gr4 = rep(0,ncol(Y))
> gr3 = matrix(0,nrow=ncol(X),ncol=ncol(Y))
> gr1.b = matrix(0,nrow=nrow(X),ncol=ncol(X))
> gr2.b = rep(0,nrow(X))
> eta = matrix(0,nrow=nrow(X),ncol=ncol(Y))
> d.eta.3 = array(0,dim=c(nrow(X),ncol(X),ncol(Y)))
> d.eta.4 = matrix(0,nrow=nrow(X),ncol=ncol(Y))
> gr3.b1 = array(0,dim=c(nrow(X),ncol(X),ncol(Y)))
> gr4.b1 = matrix(0,nrow=nrow(X),ncol=ncol(Y))
>
> #Gradiant of alpha and beta
>
> for (i in 1:nrow(X))
> {
> gr1.b[i,] = (2*p_tilde[i]-1)*((exp(-t(x1)%*%X[i,]-x2)*X[i,])/(1+exp(-t(x1)%*%X[i,]-x2))^2)
> gr2.b[i] = (2*p_tilde[i]-1)*((exp(-t(x1)%*%X[i,]-x2))/(1+exp(-t(x1)%*%X[i,]-x2))^2)
> }
> for (j in 1:ncol(X))
> {
> gr1[j] = sum(gr1.b[,j])
> }
> gr2 = sum(gr2.b)
>
>
> #Gradiant de w et gamma
> for (i in 1:nrow(X))
> {
> for (t in 1:ncol(Y))
> {
> eta[i,t] = 1/(1+exp(-t(x3[,t])%*%X[i,]-x4[t]))
> d.eta.3[i,,t] = eta[i,t]*(1-eta[i,t])*X[i,]
> d.eta.4[i,t] = eta[i,t]*(1-eta[i,t])
> gr3.b1[i,,t] = p_tilde[i]*((-abs(Y[i,t]-Yb[i]))*(1-eta[i,t])^(-1)+(1-abs(Y[i,t]-Yb[i]))*
> (eta[i,t])^(-1))*d.eta.3[i,,t]
> gr4.b1[i,t] = p_tilde[i]*((-abs(Y[i,t]-Yb[i]))*(1-eta[i,t])^(-1)+(1-abs(Y[i,t]-Yb[i]))*
> (eta[i,t])^(-1))*d.eta.4[i,t]
> }
> }
> for (t in 1:ncol(Y))
> {
> for (j in 1:ncol(X))
> {
> gr3[j,t] = sum(gr3.b1[,j,t])
> }
> gr4[t] = sum(gr4.b1[,t])
> }
> c(-gr1,-gr2,-gr3,-gr4)
> }
>
> opt = optim(c(alpha[,c+1],beta[c+1],w,gamma),g,gr,method="BFGS")
>
> The problem is that it gives me wrong results, and I have noticed that if I change my function to maximize (for example if, instead of -(ncol(Y)*sum(res3)+sum(res3.b)), I try to maximise -(ncol(Y)*sum(res3)), it gives me the exactly same results...which is not possible!
> So maybe I am using optim() in a wrong way...Does someone have an idea what could be wrong in my code ?
>
> Thank you very much in advance
>        [[alternative HTML version deleted]]
>
> ______________________________________________
> R-help at r-project.org 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.
>



-- 
===============================================
Jon Daily
Technician
===============================================
#!/usr/bin/env outside
# It's great, trust me.