more NLP (local nonconvex) primitives

Answered

Frederik Geth

September 09, 2026 01:48

I'm doing local, nonconvex NLP (for AC OPF in disribution networks) where I also want to include some more free-form NLP constraints using JuMP's user-defined functions (smooth). Specifically, I use log1pexp calls through StatsFuns.jl, as I explain here: https://frederikgeth.github.io/BMOPFTools.jl/dev/relu_softplus_encoding/
 

The naive nested evaluation of log and exp is obviously not very numerically stable

 

Gurobi does not support log1pexp or softplus primitives it appears (https://github.com/jump-dev/Gurobi.jl/blob/39e019858e3c1abda5c26bab09fc729a21bb663a/src/MOI_wrapper/MOI_nonlinear.jl#L11-L30), so I used logistic instead. I can confirm that runs - but it proves to not be as nicely behaved numerically as softplus.
 

Cheers,

Fred
 

Related to

• Nonlinear

0

• Official comment

  

  Hassan Hijazi

  • Gurobi Staff

  September 09, 2026 15:54

  Thanks Fred, we are considering supporting Softplus for v15, stay tuned.

Please sign in to leave a comment.