Bug appears when doing two-dimensional maximization

[Stan125]

Hi @alexanderlange53 and @kaihusmann,

I found a bug that appears when doing two-dimensional log-likelihood maximisation:

library("optimization")
#> Loading required package: Rcpp
library("gamlss.dist")
#> Loading required package: MASS
y <- rWARING(100)
ll <- function(par, y) {
  mu <- par[1]
  sigma <- par[2]
  if (mu <= 0 | sigma <= 0)
    return(NA)
  l_i <- dWARING(x = y, mu = mu, sigma = sigma, log = TRUE)
  s_li <- sum(l_i)
  return(s_li)
}
ll_no_y <- function(par) {
  return(ll(par, y = y))
}
optimization::optim_nm(fun = ll_no_y, k = 2, start = c(0.5, 0.5),
                       trace = TRUE, maximum = TRUE)
#> Error in if (Ry > simplex[k, 1]) {: missing value where TRUE/FALSE needed

and wanted to let you guys know. If I find the time I'll look into it myself!

Best,
Stani

[Stan125]

Sent a pull request with a solution

[Stan125]

Ok, so actually my PR doesn't solve the issue. The issue arises when the function that is supposed to be optimized is actually constricted in terms of the parameter space, and sometimes returns NA values:

library("optimization")
#> Loading required package: Rcpp
library("gamlss.dist")
#> Loading required package: MASS
set.seed(123)
y <- rWARING(100)
ll <- function(par, y) {
  mu <- par[1]
  sigma <- par[2]
  if (mu <= 0 | sigma <= 0)
    return(NA)
  l_i <- dWARING(x = y, mu = mu, sigma = sigma, log = TRUE)
  s_li <- sum(l_i)
  return(s_li)
}
ll_no_y <- function(par) {
  return(ll(par, y = y))
}
optimization::optim_nm(fun = ll_no_y, k = 2, start = c(0.5, 0.5),
                       trace = TRUE, maximum = TRUE)
#> Error in if (Ry > simplex[k, 1]) {: missing value where TRUE/FALSE needed