Replacing a sigmoid output with a Gurobi-friendly bounded activation

Awaiting user input

Ehsan

September 22, 2025 16:35

• 

I’m embedding a PyTorch MLP into Gurobi using gurobi-ml’s add_predictor_constr. My network originally used a sigmoid on the last layer to keep the prediction in [0,1]. Since gurobi-ml doesn’t support sigmoid/tanh, I’m looking for a solver-friendly replacement that still keeps outputs in [0,1] without post-hoc clipping.

Related to

• gurobi-ml

0

• 

  Pierre Bonami

  • Gurobi Staff

  September 23, 2025 07:26

  Hi,

  We actually have an open issue about this in our github repository https://github.com/Gurobi/gurobi-machinelearning/issues/347

  As I commented there, the mechanics of supporting sigmoid/tanh can be done (i.e. producing a model). Now what's more incertain is the size of models that can currently be solved with Gurobi. If you had an example that you can share (privately if necessary), we could try it out.

  0

Please sign in to leave a comment.