MS-Bayes?

I received the following email:

Did you know that it looks like Microsoft is entering the modeling game? I mean, outside of Excel. I recently received an email at work from a MS research contractor looking for ppl that program in R, SAS, Matlab, Excel, and Mathematica. . . . So far I [the person who sent me this email] haven’t seen anything about applying any actual models. Only stuff about assigning variables, deleting rows, merging tables, etc. I don’t know how common knowledge this all is within the statistical community. I did a quick google search for the name of the programming language and didn’t come up with anything.

That sounds cool. Working with anything from Microsoft sounds pretty horrible, but it would be useful to have another modeling language out there, just for checking our answers if nothing else.

6 thoughts on “MS-Bayes?

  1. Is this Microsoft, or Microsoft Research? I know Drew Purves at Microsoft Research in Cambridge (the original, not that upstart on the other side of the pond) is developing some Bayesian software.

  2. Andrew,

    Would you really want to use a Microsoft product to check your answers? Microsoft has a 20 year track-record of being unable to fix simple inaccuracies in Excel's statistical procedures.

    B. D. McCullough
    "Special Section on Microsoft Excel 2007"
    Computational Statistics and Data Analysis 52(10), 4568-4569, 2008

    When Microsoft programmed the ten lines of code that constitute the Wichmann-Hill random generator, they got it wrong: twice!

    B. D. McCullough
    "Microsoft Excel's 'Not the Wichmann-Hill' Random Number Generators"
    Computational Statistics and Data Analysis 52(10), 4587-4593, 2008

    Regards,

    Bruce

  3. About 20 years ago, MS acquired a Silicon Valley firm that was specializing in Bayesian networks. David Heckerman, Eric Horvitz, and some other people who are escaping my memory at the moment moved to Redmond to start a Bayesian research group. A lot of very good people have passed through that group (either as employees or interns) and they have published a lot of good stuff. They are mostly computer scientists (although I think that Heckerman and Horvitz came up through the Stanford decision analysis program), and so they tend to publish in Machine Learning and other computer science journals instead of the Stat ones. There is a lot of interesting stuff about Bayesian mixture modeling and other things.

    The quoted passage reminds me of MSBNx, software for pushing probabilities around Bayesian networks that MS acquired in the purchase. The marketing department wasn't used to working with such small volume projects, so they gave it away for teaching and research projects. It was once widely used, but as no real effort went into maintenance, I tend to steer students towards other things instead.

    Two applications of Bayesian thinking that have made its way into MS products that I remember. One, they had a highly successful expert system for debugging printer problems that was based on a Bayes net. The second is the paper clip. The original idea, as Eric Horvitz described it was to produce a Bayesian model of user actions and then predict shortcuts that they might like to know about. Of course, once the marketing people got a hold of it, it went horribly wrong. Interesting concept though.

    Actually, they are an interesting bunch to work with as most of them have been trained in Bayesian statistics, but never trained in classical statistics. But MS Research at least has been heavily Bayesian for 20+ years.

  4. Nathaniel: I also assumed this was referring to Infer.NET.

    About two years ago when I last checked it out, it looked very powerful and well designed. Anyone on the Microsoft platform (or with Mono) who is interested in graphical networks or probabilistic programming (eg as with Avi Pfeffer's IBAL language) should definitely download and experiment with it.

Comments are closed.