remove con_outputs fro model_io

[yunlongxu-numagic]

should we get rid of con_outputs from model IO? it is only used for OCP, and really defined by the OCP, and selected from the existing model outputs. So a model developer really isn't aware of this!

What makes things a bit more complicated is that for efficiency, the compiled model will only have this 'con_outputs' as outputs, and with derivatives taken on it. So for the compiled model, it does need to be aware of con_outputs!

Some guiding principles:

• someone who writes a model really shouldn't care what outputs are used as constraints
• the model only cares about its execution, it's the OCP that should care about what NLP functions the model provides to the OCP
• at the same time, we'd want to be able to change to a model written in a different software, but can be executed in python, and run OCP with it, and that should only change the 'model'
• ideally one should be able to change which outputs to put constraints on, even after a model is constructed
• clean up the logistics on what indices do the model and the ocp need to manage respectively
• we would still like to keep the OCP agnostic of the backend of choice and hence all the transformations used

Some challenges:

• con_outputs names are used quite widely, and deeply, to construct variable indices, orders, flat inputs for the implicit functions, etc
• it is also used for construction model algebra jacobian structures
• if we create indices for array variables which also have names, then we can't use arbitrary subsets every time! Although for dict inputs and outputs, the indices and orders are less important
• as we start to decouple ocp indexing and model indexing, we need to remove and avoid using model indexing inside OCP
• information management of constraints composition of the stage_con is definitely a week point that has hard-coded assumptions at lots of places=

Some questions:

• should we keep con_outputs somewhere, or just make it a variant instance of outputs?
• who should be the owner of the implicit method and its associated derivatives? The model owns it now, but would require inputs from OCP to determine what outputs to use as constraint outputs. The OCP could also own it, but in that case, how do we bring in external model?

[yunlongxu-numagic]

IO definition

StateSpaceModel

forward

states, inputs -> states_dot, outputs

forward_with_array

states, input -> states_dot, output, outputs might be just a subset, as it is much cheaper to collect the subset before the outputs array is formed than after

implicit

states, inputs, states_dot, con_outputs -> residuals

flat implicit

as implicit, but with a flat array as inputs