Define most basic abstract `Approximation` class

aatmdelissen commented 3 years ago

This is to define the interface for a more abstract Approximation class. The current one has a lot of redundancies. We need to determine what are the most essential characteristics and requirements, shared by all Approximation's children.

Lower and upper bounds are optional (default xmin=0.0 and xmax=1?), can both be given as a float, or a np.array.
Problem sizes don't have to be given in the constructor, they can be deduced when build_sub_prob is called.
Changing the problem size (either m or n from one call to build_sub_prob to another should be accepted (?)
I don't think a move limit is part of the approximation?

Data storage (attributes

Storage of data, is by default x, g, dg and possibly ddg (numpy referenced array, so no copy), used for the last call to build_sub_prob
P matrix for intervening variables
Zero order terms
Any properties (convexity, separability)? Can these be set as class attributes? For what purpose do we need them?

Did I miss anything?

MaxvdKolk commented 3 years ago

Seems to make sense to me. Lets try to get the base one as little as possible indeed. Nothing comes to mind at this point that might be missing in the list.

Two thoughts:

P is relatively sparse, right? We might consider one of the scipy.sparse matrices
For common attributes it is nice to store them in the class directly I think. If there will be many (unsure if this is actually the case) then we could store them in some sort of collection, dict/list/...

aatmdelissen commented 3 years ago

Yes, those common attributes can nicely be defined as class attributes, since they are the same for all instances of a certain approximation class. What are class attributes?

MaxvdKolk commented 3 years ago

Nice, haven't seen that one before, but looks handy for attributes

Giannis1993 commented 3 years ago

Seems to make sense to me. Lets try to get the base one as little as possible indeed. Nothing comes to mind at this point that might be missing in the list.

Two thoughts:

P is relatively sparse, right? We might consider one of the scipy.sparse matrices

For common attributes it is nice to store them in the class directly I think. If there will be many (unsure if this is actually the case) then we could store them in some sort of collection, dict/list/...

The way it is now, P is not sparse, because it includes both the P and Q matrices of Svanberg (i.e. ourP = svanP + svanQ).

Giannis1993 commented 3 years ago

Done

artofscience / SAOR

Define most basic abstract `Approximation` class #24