[R] About error: L-BFGS-B needs finite values of 'fn'

[Deniz OZONUR]

Hi,

I am trying to obtain power of Likelihood ratio test for comparing gamma distribution against generalized gamma distribution. And so I need maximum likelihood estimates of Generalized gamma distribution with three parameters. I wrote code as follows.

 require(bbmle)
 library("bbmle")

require(flexsurv)
library("flexsurv")

 sig=0.05
 den=1000
 n=30
 apar=2 ###alpha
 bpar=3 ##beta
 cpar=2 ##c parameter


 LRatio=function(den,n,par=c(cpar,apar,bpar)){

 LR2=rep(0,den)

 count=rep(0,den)

 cpar=par[1]
 apar=par[2]
 bpar=par[3]

 for(i in 1:den){

 y=rgengamma.orig(n,shape=cpar,scale=bpar,k=apar)

 gamma4 = function(shape, scale) {
  -sum(dgamma(y, shape = shape, scale = scale,log = TRUE))
 }

 gm = mean(y)
 cv = var(y)/mean(y)

 m5 = mle2(gamma4, start = list(shape = gm/cv, scale = cv),method = "L-BFGS-B", lower =c(.00001,.00001),upper = c(Inf,Inf))


 gengamma3 = function(shape, scale,k) {
 -sum(dgengamma.orig(y, shape = shape, scale = scale,k=k,log =TRUE))
 }

 ci=mean(y)            #c initial value
 a1=ci*mean(y)^(ci-1)
 a2=ci*(ci-1)*(mean(y)^(ci-1))/2
 mu1=mean(y)^ci+a2*mean(y^2)
 mu2=(a1^2)*mean(y^2)+2*a1*a2*mean(y^3)+(a2^2)*(mean(y^4)-mean(y^2)^2)
 alp =(mu1^2)/mu2                            #alpha initial value
 bet=mean(y)*gamma(alp)/gamma(alp+(1/ci))       #beta initial value

 m6 = mle2(gengamma3, start = list(shape = ci, scale = bet, k=alp),method = "L-BFGS-B", lower = c(.00001,.00001,.00001),upper = c(Inf, Inf, nf))


 LR2[i]=2*(logLik(m6)-logLik(m5))
 count[i]=LR2[i]>=qchisq(1-sig, df=1)

 }

 pow=sum(count)/den
 print(i)
 print(pow)
 }

 But I got error : optim(par = c(3.88907163215354, 3.62005456122935, 1.66499331462506 :  L-BFGS-B needs finite values of 'fn'


What is wrong? Can you hep me, thanks..
Deniz...