Snobbing around
Snob is the grandfather of MML software. I was looking for tools for Gamma mixture modeling and was lead to Snob by Yudi Agusta and David Dowe’s (2003) work on Gamma distribution mixtures. From David Dowe’s Snob page:
Snob uses the Minimum Message Length (MML) principle to do mixture modelling to infer MML finite mixture models.
Mixture modelling (or mixture modeling) concerns modelling a statistical distribution by a mixture (or weighted sum) of other distributions. Mixture modelling is also known asunsupervised concept learning or unsupervised learning (in Artificial Intelligence) intrinsic classification (in Philosophy), or, classification clustering numerical taxonomy Minimum Message Length (MML) is a method of machine learning, statistical inference, inductive learning, "knowledge discovery" and "data mining" very much in line with the notions of Kolmogorov complexity and algorithmic information theory pioneered by R. J. Solomonoff, A. N. Kolmogorov and Greg Chaitin. See also Wallace & Dowe (1999a), "Minimum Message Length and Kolmogorov complexity", Comp. J., Vol 42, No. 4, 270-283 [which is the Computer Journal’s most downloaded “full text as .pdf'’ article - see, e.g., here].
The original Snob paper was : Wallace, C.S. and Boulton, D.M. (1968), `An Information Measure for Classification‘, Computer Journal, Vol. 11, No. 2, 1968, pp. 185-194. This is the same paper in which MML was developed. (See also more recent Snob theory and application papers.)
Snob currently deals with finite mixture models (or a finite mixture model) of
Normal (or Gaussian) distributions discrete multi-state (also called Bernoulli or categorical) distributions Poisson distributions von Mises circular distributions missing data Chris Wallace extended Snob in 1998 to deal with spatial correlation (and Markov fields), as occurs in images.
Gerhard Visser and David Dowe (2007) endeavoured to extend Wallace (1998, above) with "Minimum Message Length Clustering Of Spatially-Correlated Data with Varying Inter-Class Penalties".Russell Edwards and David Dowe created a version of Snob (also in 1998) which deals with single Gaussian factor analysis in sequentially and spatially uncorrelated data. It uses total assignment. (See publications.) Yudi Agusta and David Dowe have also developed MML mixture modelling software for other correlated Gaussians, t distributions (2002) and Gamma distributions (2003, .pdf); and Jon Oliver and David Dowe published a note (1996) on MML mixture modelling of von Mises-Fisher spherical distributions.
The age of the program shows. The code is still Fortran 77 and is console only. I downloaded GNU Fortran, installed in C:\gfortran rathet than "Program files …" for fear that it may not recognize long file name and spaces in the file path.
Then I tried to compile the source code. I have never written Fortran and the last time I read it was some 20 years ago. There is a Make file for Linux. I couldn’t find how to use Make under Windows at GFortranGettingStarted - GCC Wiki, so I tried the dumbest thing:
gfortran *.f
Dumb luck. After pages of warning, the compiler halted at 3 errors. The last was a complaint about "fdate()" undefined. It was called from "date.f". A quick google check confirmed that it’s the same as CTIME(TIME8()). I substituted it in "date.f" and compiled again. Down to 2 errors.
The first error was about "bdgBsf()" defined twice. It indeed was defined in 2 files. "A_lgI0.f" appeared to be obsolete, so I renamed it as "A-lgI0.f_old" and recompiled. Vola, I got "a.exe" despite all the warnings. Apparently "a.exe" runs under DOS with no problem.
I should mention that I stumbled up on two heirs of Snob. Little Snob is now obsolete, suppassed by PyMML, a Python package. Apparently they are from the same university, Monash University, Australia, where Snob was created. I emailed David Dowe to see if there is any update.
Still have to read the 35 page user’s manual, written in 1996.