Henning Thielemann
On Sun, 12 Sep 2021, Dominik Schrempf wrote:
I work with Markov chains (yes, I had to write my own MCMC library in order to run proper Markov chains),
I don't know what proper Markov chains are, at least I wrote my own simple lazy Markov chain generator years ago that fulfilled my needs: https://hackage.haskell.org/package/markov-chain
I have also written a package for Hidden Markov models: https://hackage.haskell.org/package/hmm-lapack
Thank you for mentioning this. I was playing around with 'markov-chain' some time ago! I was referring for something more generic, for example, where the transition rate/probability matrices can be directly defined. I am not so familiar with hidden Markov models. With respect to the term "proper", thanks for picking that up :); I should have been more specific: I was referring to Markov chain Monte Carlo samplers. For a reference implementation, please see the 'mcmc' library in R [1], for a feature-rich one, e.g., 'PyMC3' [2].
I need to plot my data (there is no superb standard plotting library available in Haskell). By now, I do maintain library packages providing answers to some of these problems, but it was (and is) a lot of work.
If you write packages that fulfill the needs of your applications, that's certainly better than writing impressive frameworks where no one knows whether it is usable, at all. :-)
You are right, we want to avoid impressive but unusable frameworks. I was referring to the following question: "For a given goal, which features do I have to implement myself in order achieve the goal?" With the existing set of Haskell libraries, I make the following observation too often (or too early in my stack of requirements): "This feature is not available (or not readily available), I need to implement it myself." For example, at the moment I am working on a library for estimating covariance matrices from sample data. This is not at all my area of expertise. In Python and R, well maintained state-of-the-art methods are readily available (e.g., shrinkage based estimators such as the Ledoit-Wolf estimator, or oracle approximating shrinkage, as well as estimators based on Gaussian graphical models such as the graphical lasso --- actually, there is a glasso library on Hackage but it seems unmaintained, lacks documentation, and improvements from newer findings in the last 10 years). I hope I made myself clear. I don't want to naysay. I really enjoy Haskell per se! I am a big fan. Dominik [1] https://mran.microsoft.com/snapshot/2015-02-27/web/packages/mcmc/index.html [2] https://docs.pymc.io/