Accessing optimization variables during optimization

Answered

Daniel Johnson

• February 17, 2023 18:36
• Edited

(This is a repost of this post on the help centre as I was advised to ask here instead)

Hi everyone, 

I'm a beginner to Gurobi, so i apologise if this is an obvious question. 

I have an optimisation problem involving a physics simulation which I'm using Gurobi on. In order to evaluate the objective function I need to make calls to the simulator, which means I need to access the values of the constraints at that moment. 

From what I've found online it seems that generally you can't access the values of the variables until after the optimization has concluded, meaning I would be unable to make these calls to the simulator in my objective function.

I've reviewed this post (which seems to have been abandoned by the original poster) and this post, though it seems to me that neither of these generalise well to my situation. 

Is there a way around this limitation that would allow me to access these variables? 

Thanks for any help!

Related to

• General Constraints
• Formulation

1

• 

  Riley Clement

  Gurobi Staff

  • April 06, 2023 00:00
  • Edited

  Hi Daniel,

  It's certainly not an obvious question!

  We don't have the ability to embed simulation models in the optimization process, but as of Gurobi 10 we do have the ability to incorporate a different type of model from the predictive analytics space into Gurobi models - a machine learning (ML) model.

  You could take an approach called surrogate modeling, where you use a ML model to approximate a function, or indeed another model like a simulation.  Why?  Because ML models are fast to turn inputs into outputs.  The idea would be to use your simulation model to generate a very large dataset which maps input values for your objective function, to the objective function value.  You would then choose a ML model type that is both supported by gurobi-machinelearning and a good choice for your type of problem.  The ML model can then be linked with variables in your Gurobi optimization model - including a variable representing the desired value of the objective function.
  
  There are a couple of obstacles here though:

  1) Note making a good choice of ML model, creating an appropriate and balanced dataset, and training the ML model to provide a good surrogate model is not a trivial task for those not experienced in machine learning.

  2) The ML model will only approximate the simulation model/objective function.  You may be able to get a very close approximation, but unlikely to be able to replicate it with 100% fidelity.

  If you go down this route, please publish your project to a github account and share a link with us, I'm sure we'd love to see it in action.

  - Riley

   

   

   

  0
• 

  Daniel Johnson

  • February 20, 2023 11:31

  

  

  

  Hi Riley,

  Thanks so much for your response! It was very thorough and helpful! 

  I am unsure if I should go down that route at this stage as I have no experience with ML methods and this sounds like it would be a significant undertaking. However, I am glad to know that this option is on the table as it's something I may come back to in the future if I need to optimise the execution time of my algorithms.

  Thanks for your help!

  0
• 

  Riley Clement

  Gurobi Staff

  • February 20, 2023 12:08

  No worries Daniel, perhaps a simple metaheuristic such as simulated annealing would be better in this case as the objective function can be a "black box".  You may find a handy python framework if you search, but it is also not too hard to code this yourself.

  - Riley

  0

Please sign in to leave a comment.