Limitations of the internal algorithm interface

[janosg]

This is a meta issue where we collect things that are hard or impossible with the current internal algorithm interface. If we collect enough issues, we hopefully see similarities and can address them. Most issues we know so far come from algorithm specific options that are related to the params (e.g. an initial simplex or an initial population). Users should be able to specify those in the external parameter space the algorithms need them in a flat array format.

[sefmef]

For the following algorithms the corresponding options are missing:

• NelderMead - initial_simplex
• BFGS - hess_inv
• Powell - direc

[janosg]

Genetic optimizers or multistart could benefit from providing an initial population or set of starting points. The simplest solution for supporting this would be to expose the converter as part of the InternalOptimizationProblem (it's already there as a private attribute) and use it in the algortihms to convert user provided parameter vectors to flat internal representations.

[spline2hg]

@janosg I would like to work on this, as its also needed for the PYGAD GA.

[janosg]

Nice, please go ahead. I texted you some details on zulip. It will be a small change.

[janosg]

Another limitation: We currently don't have a way to mark algo options that need a sign switch for maximizatio. Examples would be stopping_funvalue in GFO.