Like the lottery... (Migrated from community.research.microsoft.com) RRS feed

  • Question

  • JimGale posted on 01-27-2010 7:38 PM

    Hi. First, thanks for writing Infer.NET - I look forward to using it correctly.

    Second, although it would be an interesting experiment, I'm not really writing/researching anything to do with the lottery. [this is nonprofit after all].  However, it makes for a very good comparison and simplifies my questions...

    Imagine I wish to observe a correlated set of numbers and the amount people bet on those numbers, or win from those numbers [using the lottery analogy].   I wish to plug in other numbers later to see, statistically, what amount would be most relevant.

    As an example:
    Observation 1: 1-5-10-12-15-16 nets 120.
    Observation 2: 1-6-11-12-16-20 nets 130.
    Observation 3: 32-33-34-40-41-42 nets 400. [as in most people don't bet outside of birthday's, for example].
    ... to 1000 observations.

    One challenge here that I want to understand how to represent properly is that ALL OTHER NUMBER SETS produce/net zero, as they were not observed (i.e. didn't win / weren't drawn) - or implicitly observed as producing zero?  If they weren't 'drawn' (analogy), then they aren't meant to be together.

    *Intentially assuming* that some numbers correlate to other numbers, I want to plug in a set of numbers to see how likely they would produce anything relevant.  I see the point machine example, however, it appears to house only logical true/false observations.  My attempts to change it to produce a value only cause a "specific output is not supported yet" -like message.  Also, do I need to randomly add lots of other possibilities that match an implicit observable of zero?  [because, much like the lottery example, that may be many many zero-producing combinations].

    Or, is there a better way for me to:
    1. Add only my observables in order to perform valid tests?
    or 2. Even better would be to see which numbers correlate to other numbers in my tests.  Analogy: Would 1-2-3 be very likely to be matched with 20-25-31?  Is there a good approach to that using Infer.NET - and an example?

    I hope my analogy wasn't too distracting and allowed the questions to be clear enough,

    Thanks again,

    Jim Gale

    Friday, June 3, 2011 5:33 PM


  • minka replied on 01-28-2010 10:04 AM

    In order to use Infer.NET, you need to have a probabilistic model (possibly with unknown parameters) that describes the data.  So far you have described the data but no model.

    Friday, June 3, 2011 5:33 PM