Objective is -Inf in weighted EM

[william-denault]

Hello,

I am using mixsqp to solve a weighted EM problem. I encounter some problems when some of the weights is close or equal to one. The problem persist even when I change some of the control values.

load("problem_weighted_mixsqp.RData")
L_pb <- dat$L_mat
w <-  dat$zeta
tlength <- dim(L_pb )[2]-1
mixsqp_out <- mixsqp( L_pb ,
                     w,
                      x0 = c(1, rep(1e-30,  tlength )),
                      log=TRUE ,
                      control  =  list(
                        eps = 1e-6,
                        numiter.em = 40,
                        verbose = FALSE
                      )
)

problem_weighted_mixsqp.zip

[pcarbo]

@william-denault Try this:

load("problem_weighted_mixsqp.RData")
L_pb <- dat$L_mat
w <- dat$zeta
tlength <- ncol(L_pb) - 1
out <- mixsqp(L_pb,w,x0 = c(1,rep(0.001,tlength)),log = TRUE,
              control = list(eps = 1e-6,numiter.em = 4,
                       	     verbose = TRUE))

Starting too close to the boundary of the feasible set can sometimes cause numerical issues.