Summary of issues

[mlprt]

Structural

I have some uncertainty about the large-scale design of Feedbax. Are there better abstractions we should use?

In approximate order of significance:

1. General design concerns with `AbstractStagedModel` and `AbstractIntervenor` #28
2. ModelInput and passing intervenor parameters to submodels #12
3. Structure of mechanical models: effector states, and aggregation of dynamic and static updates #29
4. Intervenors, and an issue with the abstract-final pattern #20
5. Intervening on model parameters #21
6. Training, and vmapping over ensembles of models #5
7. Multiple feedback channels and MultiModel #30
8. Does the model need access to the task, at construction time?  #16
9. Associate types of intervenors with types of staged models #17
10. Eliminate AbstractTaskTrialSpec #35 and The suitability of WhereDict: lambdas as keys #14

Typing

I'm currently working to make Feedbax compliant with the Pyright static type checker. The following issues involve typing, but not on such a scale that I predict they would lead to large-scale structural changes.

• The abstract-final pattern and generics: should AbstractState be eliminated? #24
• Typing ModelStage #23
• Support Python>=3.9 #25
• Jaxtyping and PyTrees with mixed array and non-array leaves #8
• Typing of scalar derivatives of state variables #11

Features

• Intervening on model parameters #21
• Improve the realism and generality of models of musculoskeletal geometry #33
• Include any neural network as a component of a staged model #2
• Generalized loss terms from trial specifications #10

Performance

TODO.