Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Interesting result from SSDF

Author: Bruce Moreland

Date: 14:36:08 06/08/00

Go up one level in this thread


On June 08, 2000 at 15:25:22, Amir Ban wrote:

>On June 07, 2000 at 19:05:30, Bruce Moreland wrote:
>
>>On June 06, 2000 at 15:53:44, Wayne Lowrance wrote:
>>
>>>All of the probobability details  boil down to Junior having about as much
>>>chance of beating Crafty 10-0 as I do winning the lotto, ok, exagerated but the
>>>result from my perspective is wrong, something is wrong, I agree with Dr Bob's
>>>assessmenmt. I am a educated EE (retired)and If one of my staff engineers came
>>>to me with a similiar test result of his project, well, he would know better not
>>>to come to me with it, Period.
>>
>>You would be being too harsh.
>>
>>It's true that if you pick a random string of results, you shouldn't expect to
>>see an improbable outcome.
>>
>>But if you can search through a series of sub-strings, looking for one that has
>>an improbable outcome, you might very well find one, since you give yourself a
>>lot of chances.
>>
>>I wrote a program that allowed me to test this.  I did a series of matches.
>>Each match was 30 tests.  If a 60/40 probability came back on the 60 side, I
>>scored a given test a success.  If there was a string of 10 successes in a row
>>in any match, I counted this match.  It should have been very possible to
>>predict the probabilities here exactly, but I figured a Monte Carlo simulation
>>would be good enough.
>>
>>I ran 1,000,000 trials and 55,000 of them had a series of 10 successes in a row.
>> Which is 5% of them, which sounds like not a lot, but it is really quite a lot
>>given how rare a string of 10 successes in a row should be.
>>
>>If I use a 50/50 probability, I only get about 10,000 matches counted, and if I
>>up it to 75/25, I get over 300,000, which is 30% of the matches.  It seems like
>>small differences in the probabilities used have a large effect upon the
>>outcome.
>>
>>This doesn't simulate chess very well, since we have draws, but I think that
>>I've shown the test shouldn't be summarily dismissed as unfair.  It may have
>>been an improbable outcome, but not necessarily so improbable that you'd suspect
>>the test.
>>
>>The only thing I'm a little suspicious of is that the string was the last ten
>>games, so I'd wonder if Crafty was a little sick or something.  That could bear
>>some investigating.
>>
>>bruce
>
>Ok, here's my two cents:
>
>The most likely estimate for the game outcome is the overall match result
>(33.5/41). The probability for any *particular* sequence of 10 games (say the
>last 10) to be 10-0 is (33.5/41)^10 = 13.3%.
>
>If you want probability for any sequence anywhere, the probability is larger, as
>you point out.
>
>If you suspect the 10-0 result, you can argue that the most likely estimate
>should discount this result, so this gives you a probability of (23.5/31)^10 =
>6.3%. Again, much larger if you look at *any* sequence. There's not much to say
>for this argument, because it discounts the 10 games on the premise that the
>result is improbable, but still finds that the result is probable.
>
>Amir

As Uri pointed out, it's hard to relate this to chess, since there are draws.
If you score 75% against someone, that isn't enough information.

If you score 75% by winning three out of four, the odds of 10 wins in a row are
(3/4)^10, which is 5.6%

If you score 75% by winning half the games and drawing the rest, your odds of 10
wins in a row are (1/2)^10, which is only 0.10%.  That's a dramatic difference.

My guess is that ten wins in a row, given an evident ability to score about 75%,
isn't so far out of the question that I'd doubt the test.  That was the point I
was suggesting to the fellow I responded to.

bruce



This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.