Running a Bi-level scenario in HPC where the outer design space variable is days of the year and the sub scenario variables are some angles u and beta. I am trying to first run a coarse full factorial DOE in the sub scenario to find semi-optimal point for each day which will then be used as an initial guess for the local gradient based optimizer.
I need a way to pass the initial guess to either the same sub scenario (but change the algorithm from DOE to local optimiser) or a new subsenario, after the coarse DOE is complete. This needs to happen mid-execution of the outer scenario. And I need to be able to specify the variable value for each day. Any suggestions will be appreciated.
The issue with multi-start is that the custom doe algorithm requires samples which are then used as the initial points for the optimisation algorithm.
We would like these initial points to be the optimalresult (design variables: u, beta) of a DOE scenario’s execution, which are not known before execution.
To be more clear, the multi start algorithm would process and run the optimiser for all samples, whereas we only want to run the optimiser for the most optimal sample)
Ok that’s clear. For this we need to develop a specific driver that combines a DOE and an optimizer, and uses the DOE as starting point. It is easy with a single level scenario (just execute the scenario first with the DOE and then with the optimizer), but for an inner level scenario it would not work. The issue that Jean-CHristophe create would handle that.
Following our discussion with @MatthiasDeLozzo and @Francois.Gallard, we could easily craft an MR. I’ll bootstrap that quickly (hopefully this week) and put it in review. It would be in GEMSEO v7 but the logic would be somehow lightweight and you could copy and paste it in your project in the meantime.