POEM_096: Minimization of Constraint Violation

[andrewellis55]

No description provided.

[robfalck]

Thanks for initiating this.

I have a couple of questions about implementation.

1. Is the constraint violation just the L2 norm of the driver-scaled constraints? Is there any reason to support any other ways of computing it?

2. From an API standpoint, should feasible_point=True be an argument to run_driver, or an option? I'm somewhat in favor of it being a run_driver argument. Since we're redefining the optimization problem after setup there may be issues with relevance

This would further boost the case for refactoring the drivers into OptimizationDrivers and AnalysisDrivers, since this wouldn't make sense in the latter context.

[andrewellis55]

1. I believe that SNOPT uses the L1 norm to find a feasible point

...

I don't recall Dr. Martin's textbook mentioning minimization of constraint violation, but where he mentions merit functions it's suggested to use either L1 or L2

I'm not an expert on the topic but perhaps some of the academic collaborators may have a suggestion\preference.

2. I agree with it being a input to run_driver. I think feasible_point could be a parameter name, although I could imagine some other option might be find_feasible_point, feasibility_optimization or minimize_constraint_violation so that the parameter name has a stronger implication that it's changing what the driver it doing (from minimizing the objective to finding a feasible point with no regard for the objective).