AI / inference problem

Question

Let's say I have 20 players [names A .. T] in a tournament. The rules of the tournament state that each player plays every other player twice [A vs B, B vs A, A vs C .. etc]. With 20 players, there will be a total of 380 matches.

In each match, there are three possible outcomes - player 1 wins, player 2 wins, or draw. There's a betting exchange which, ahead of each match, quotes the probabilites of each outcome occuring; so you might have 40% player 1 wins, 30% player 2 wins, 30% draw [probabilities sum to 100%]; I store these probabilities ahead of each match.

Fast forward one quarter of the way through the tournament. I have collected probabilities for 95 games, with 285 still to go. What I want to know is -

Can the probability data from the 95 games be used to predict probabilities for the remaining 285 ?

For example, if I know A vs B and B vs C, can I use them to infer A vs C ?

And if so, how do I do it ?

Answer 1

Let me introduce you to my good friend Bayes... http://en.wikipedia.org/wiki/Bayesian_inference

Edit: Part 1) Bayes will only work for non-independent trials. If winning one game somehow increases your probability of winning the next, you can carry on! Otherwise this isn't very helpful at all.

Edit: Part 2) Regardless, the base is the following Bayes' Formula.

P(A|B) = P(B|A) P(A)
         -----------
             P(B)

Which is read, "The probability of A given B is equal to Prob. B given A times Prob of A all over Prob. of B". To illustrate this, the car salesman with 3 doors problem is often given.

You have 3 doors and behind one door there's a brand new car. The other two doors have absolutely nothing. The host then asks you to pick a door. Remember, there is door 'A', 'B' and 'C'. Therefore, you have a 1/3 probability of being correct.

The host, being a generous guy, opens one of the other doors. He now gives you the option of either sticking with the same door or opening the other door.

I realized that explaining this in a Stackoverflow reply would take forever and just googled it. This is the Monty Hall problem: http://en.wikipedia.org/wiki/Monty_Hall_problem. http://en.wikipedia.org/wiki/Monty_Hall_problem#Bayesian_analysis for the Bayes section.

Edit: Part 3) You may want to look up 'Bayesian Networks' if you decide this sort of approach can work (but on a much grander scheme)