POEM_094.md: Driver Autoscaling and Refactor

[robfalck]

Adding POEM 094 which proposes

• Separating analysis and optimization drivers
• Providing methods to get certain optimizer information from the driver in an optimizer-independent way
• Adding an Autoscaler class to provide automatic scaling of design variables and responses.
• Implementation of one or more Autoscaling algorithms initially.

[tuckerbabcock]

I think these changes will be great an allow more flexibility from external users to write their own Drivers, and the addition of an Autoscaler class makes a lot of sense to allow more flexible scaling approaches.

About the new Autoscaler class and API, I wonder if the proposed scale method provides enough granularity for very general autoscaling algorithms that may include an automatic re-parameterization of design variables or responses like the method introduced in this paper.

I may be wrong, but I believe that the scaling capabilities that exist in OpenMDAO today are equivalent to multiplying the design variable or response vector by a diagonal matrix and adding a bias vector (with the entries for scaler on the diagonals and adder in the bias vector). More generally, at least for the design variables, I think we may want to allow for general transformation matrices that effectively re-parameterize the optimization problem, and I think this may require more internal changes to how OpenMDAO handles scaling.

I'm very excited for these changes to make their way into OpenMDAO.