Abandoning "Mamba.Chain"

[xukai92]

Current use of "Mamba.Chain"leads to problem like #390 and #406. Some of them is very hard to resolve (e.g. #390) within "Mamba.Chain"

For this reason, @yebai and I decided to build our own Chain type now. (FYI the main reason we used Mamba.Chain was to re-use all the diagnosing function written in Mamba). This would requires use to generally build these functions (like ESS) for our own chain type (or somehow reuse them in a way which allow missing values).

For describe function is self, the most important two to implement are

• ESS
• MCSE

And for plot, we may use Plot.jl to plot trace and density (in stead of Gadfly.jl).

And don't remember these functions to diagnose convergence in Mamba.jl which are also sort of wanted.

Discussion are encouraged on this!

[keorn]

Sounds like a good change, it could enable the support of arbitrary sample types too.

[xukai92]

Note: fetch matrix indexing (i.e. v[:,i]) from Mamba.Chain is broken now. Need to make sure new solution support it.

[yebai]

I just noticed that there is already a GSoC project working on the varying dimension constraint of Mamba. According to a blog from the main developer, the refactored Chain (see here) can already store samples with varying sizes. Perhaps we can adapt his code for our purpose?

[xukai92]

Another choice: arviz-devs/arviz#97 (comment)

[yebai]

Closed by new MCMCChain package.