Roadmap to public release.

[AdrienCorenflos]

This will be populated during the talk but feel free to list a number of items you think are critical to the release.

[AdrienCorenflos]

Personal critical items

I will update this list when I have another idea.

Particle filtering and smoothing

• The blueprint from state-space model formulation to Monte Carlo estimators should be absolutely clear
• The most basic version of cSMC pMH should be supported in almost the same steps as for the basic PF/PS

Interface

• The same interface should be used for both Kalman (exKF/KS) as for PF/PS: define elements and operator, and then scan over it.
• Delegation of responsibility: as it stands I do not like that inferential loops are in cuthbertlib and not cuthbert. The cuthbertlib should be about what can be shared: I can't see a world where cuthbert.pf would need to run a full KS (I'm lying, I can see a reason, but it is so niche it doesn't make sense interface wise).

Examples

• For me, at minima we need a Gaussian filter, an iterated extended smoother, a particle filter, and a cSMC.
• Parameter identification if possible in the Gaussian case (EM vs gradient MLE?) other ones can wait.

Documentation

We need to make it really clear what the diff to dynamax is: we go lower level, and share interface at the operator stage, rather than at the model definition stage + there is a big focus on particle filtering and other "discretization" methods such as embedded HMM, factorial approximations, etc.

In some sense, I would like for Nicolas to be able to write the V2 of his book using our library rather than his, no can do with dynamax.

[SamDuffield]

Agree on all of the above! Except I'd probably be happy to release without iterated smoothing methods (with a view to adding them post-release, ofc in the best case it's just a simple example and easily to support with the current code then that's totally fine)

You can add your desired examples to #38

[SamDuffield]

For me I'd add

• Examples listed in Add examples #38
• Discrete state space inference (with associative scan), which should be fairly easy
• Easy support for irregular time spaced observations (and an example showing it)
• Support for offline filtering between observation times (and an example showing it) - can be supported by observations filled with NaNs
• Particle filtering with proposal beyond bootstrap (e.g. locally optimal, extended Kalman)

Then for post release I'd like to add

• Factorial SSMs
• Online particle smoothing

[SamDuffield]

What generality non-Markovian SSM to support?

$G(x_k)$?
$G(x_{k-1:k})$?
$G(x_{1:k})$?

@AdrienCorenflos says
$G(x_{k-1:k})$ for cuthbert (for public release)
$G(x_{1:k})$ for cuthbertlib (can be post-public if need)

[SamDuffield]

• Resolve Change convention to include observation at t=0 #80
• Resolve Should we enforce model_inputs to be an attribute of all filter and smoother states? #77
• Particle smoother
• Discrete (associative)
• Docs + clean up README (statement of purpose, CONTIRBUTING.md)
• Examples see Add examples #38

[SamDuffield]

Sam prios:

• Merge Add associative filters for taylor and moments #110
• Establish format for examples Set up examples in docs as runnable python scripts #105
• Add quickstart example with international football ratings

Sahel prios:

• Discrete cuthbertlib atom
• Discrete cuthbert interface
• Add automatic docs generation (mkdocstrings)

Adrien prios:

• Merge Add associative filters for taylor and moments #110
• Example from dynamax
• Iterated Kalman example (building on quickstart football example)

[SamDuffield]

We should also make sure to address the following issues (mostly easy)

• Fix tests/cuthbert/gaussian/benchmark.py #87
• cuthbertlib.stats.multivariate_normal.logpdf fails when x is an array but mean is a float #112
• linearized_kalman_filter_state_dummy_elem does not belong in cuthbert/gaussian/types.py #115
• Improve aesthetic of automated docs rendering #120 [SD]
• Make sure latex compiles correctly in the docs #121
• Future of cuthbertlib.kalman.utils.append_tree ? #66 (not essential)

[SamDuffield]

Closing as cuthbert now public!