Negative predicted reduction in tron/trunk

[tcovert]

Hi JSO team,

What should I infer about my problem and/or its suitability for use with tron/trunk if I encounter solution status equal to :neg_pred?

The problem I am solving is unconstrained, and is definitely not globally convex (I am trying to minimize the negative log likelihood of a nested logit model). lbfgs does seem to "work" in the sense that if I give it a reasonable amount of time, it finds a reasonable fit (though I've yet to run it long enough to see if it converges). Is the failure mode I am encountering with tron and trunk telling me something about the structure of my problem? That I have perhaps a bug in my log-likelihood definition? Or something else?

When I solve a simpler non-convex problem using similar data ("vanilla" logit, but with scale parameters that vary across choice sets, this heterogeneity I believe makes the problem non-convex), both tron and trunk seem to work great. So I'm scratching my head about my nested logit setup...

Thanks in advance for any advice you might have.

[tcovert]

If it helps, I'm using autodiff (Zygote for gradients, and ForwardDiff over Zygote for the hessian-vector product).

[tmigot]

Hi @tcovert ! Thanks for the issue
Is there any chance you have a reproducible example?

[tcovert]

Sorry, I realized this was a bug in my autodiff implementation (I had written an incorrect custom adjoint), not an issue with the solver. My bad.