[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums

a bias in rollouts

Posted By: Matt Cohn-Geier
Date: Friday, 2 May 2008, at 7:14 a.m.

In Response To: a bias in rollouts (Stick)

You're assuming the rollout results are wrong and overestimating from the go - which I also don't understand.

I have centuries of mathematics to back me up ;)

Look, backgammon is a game of chance. In statistics, you aren't going to get the right answer. You're going to get near the right answer with a certain margin of error (confidence interval, or Bayesian probability, or whatever).

So far so good, I think. Now take any one play. Let's say that if we roll out this one play, we are EQUALLY likely to get a result that is lower than the true equity as we are to get a result that is higher. This may be somewhat of a bastardization of Bob's original intent but I think it expresses the general sentiment.

Now take three plays which give exactly the same true equity, 0. Play A will roll out lower than the true equity with probability .5 or higher than the true equity with probability .5. Likewise for play B and C.

What's the probability that all three plays roll out to lower than 0? (1/2)^3 = 1/8. The probability that any one play is higher than 0 is 7/8. So with a probability of 87.5% the rollout result will be higher than the true equity, whereas it's only 12.5% that it's lower than the true equity. Meaning that the method we use is biased for equities higher than the true one.

You can't create a position to demonstate what you're talking about? Leads me to believe such a bias is either non existant or not worth worrying about.

The best known method we have of determining the equity of a position is rollouts. Short of that, you can calculate the equity directly, but that's extremely difficult even for simple bearoffs. You can't use rollout results to demonstrate a bias in rollouts without getting into some vicious circles. The bias should be there as a result of certain statistical principles and methods, not because I unjustifiably assumed it.

I also roll out my permanent rollouts so many trials that, as you mentioned, who gives a shit if after 15k trials it's +.5506 or +.5505.

Like I said, I'm certain that you personally don't care, but I'm not certain that it doesn't make a difference in other ways, like the odd-ply vs. even-ply mentioned on this thread. I also don't believe that there is any better method than rollouts, if that's what your issue is.

Messages In This Thread

 

Post Response

Your Name:
Your E-Mail Address:
Subject:
Message:

If necessary, enter your password below:

Password:

 

 

[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums is maintained by Stick with WebBBS 5.12.