Expectation maximization

[recidive]

trafficstars

Hello, thanks for your great library.

I was wondering how hard it would be to implement Expectation Maximization for generating the matrixes, specially the observation covariance matrix. For better filtering performance on non-linear systems. Like pykalman does here:

https://github.com/pykalman/pykalman/blob/master/pykalman/standard.py#L1339

Maybe the implementation can make use of this library:

https://github.com/lovasoa/expectation-maximization

Can you give me any guidance on how to make that?

[piercus]

@recidive I feel this is a very good idea, but it will require some work ! I can give support on this, but i won't have time in following months for doing the implementation.

Strategy

EM algorithm is based on a metric of 'how good does the KF match the observation'.

There are 2 different possible (and valid) approaches for this :

• We could reuse the idea of the README.md (measure KF model against observation) and use the mahalanobis distance for the EM algorithm
• We could do the same thing as in pykalman, and before implementing the EM, we should focus on log-likelihood (see #44)

I feel, one should worry, on the mathematical ground, about the deep difference between log likelihood and mahalanobis for us :

• It is equivalent ?
• If not what difference does it make for us ?

What do you think ?

EM Lib

expectation-maximization seems a good implementation, anyway it is depending on numeric and multivariate-gaussian which dependencies i do not have yet (i have reimplemented most of lineagebra i needed to avoid big footprint).

So i feel we could do a first implementation here, but we would need to discuss the added footprint of adding this additionnal deps in the project.

Maybe there are some workaround like :

• Implement kalman-filter-em package as a separated lib from the kalman-filter library, so we can require both lib when optimizing the KF, but only the core lib when using the KF
• Reimplement the EM code here

Anyway, i feel this question is not the top priority, and we could work with the expectation-maximization now, once unit test will pass and the results will be alright, we will take care of the final package footprint.

TDD

If you want to work on this, i would like to be able to discuss the unit test and the signatures before the beginning of the implementation.

I'd like to work in TDD (Test-Driven-Development), (A) first create unit test and make them fail (B) then code the feature

I would suggest like :

• You create a branch <feature> for the project
• You create another a branch (from <feature>) <feature>-unit-test
• We use a PR to validate the, still failing, unit test together, and discuss them
• Once test are agreed (but still failing), we merge <feature>-unit-test -> <feature>