Which of the optimization solvers in NLopt are you talking about, there are many.
I am a contributor to the 'nloptr' R package and my experience is that most/all of the solvers in NLopt are accurate and more reliable than any of the (many!) optimization functions in R. PS: Might be better to ask such wuestions on the "julia-opt" mailing list, perhaps some of the Julia optimization specialist are more inclined to answer questions there. On Thursday, May 15, 2014 12:51:20 PM UTC+2, Tim Holy wrote: > > I've had similar experiences. For me, Optim.jl is currently somewhat more > reliable than NLopt, although it's not as if Optim.jl is without its > problems. > I confess I haven't tried hard to track down what's happening with NLopt. > In > Optim, there's a branch, teh/constrained, which contains some > not-yet-worthy- > of-merging work but which (in my hands) further enhances Optim's > reliability. > If you like, you can check that branch out yourself and experiment with > it. > > Another alternative to consider is JuMP/Ipopt; I don't yet have a lot of > experience with it, but Ipopt is well-respected. > > Best, > --Tim > > On Thursday, May 15, 2014 03:23:56 AM Tom Nickson wrote: > > Hi all, > > > > I am trying to optimise the log-likelihood of the Gaussian Process. This > is > > a straight port of some code form MATLAB, so I know the gradients are > > correct. Using Optim.jl I don't have too many problems (I was one told > > "dphia < 0" however I can't replicate it). > > Using NLopt, which the documentation seems to imply should be more > stable, > > I regularly get failures if I try to run for more than a couple of > > iterations - generally it will work with 5 to 10, no more. The exit code > is > > quite unhelpful, simply "NLopt failure". I have no problems running > > non-gradient based methods, (eg COBYLA) which would lead me to think my > > gradients where incorrect - except that they work in MATLAB and with > > Optim.jl, and checkout with finite differencing. > > > > Any ideas of how to start debugging this? I could do with being able to > > apply constraints. > > > > Tom >
