Strength of program
Dragomir R. Radev
Different players play against opponents of different strength. So, two
people with the same rating, let's say 1610 rating might have had
completely different programs in terms of the strength of their opponents -
one may have played only 1200-rated players, whereas the other may have
played only 1800-rated players.
An interesting statistic would be to have for each player, the average
rating of his opponents.
Drago (radev on FIBS)
--
Dragomir R. Radev Graduate Research Assistant
Natural Language Processing Group Columbia University CS Department
Office: (212) 939-7121 Lab: (212) 939-7108 Home: (212) 749-9770
http://www.cs.columbia.edu/~radev
Martin Ronek
the past.
Never having been rated myself, but having played on several other
commercial bg servers, the fibs ratings system came as a surprise, and
initially, as a distraction to me.
It is far too easy to become preoccupied with one's ratings to the
exclusion of the enjoyment and learning of the game itself.
Much has been posted on the subject of "maximizing" one's ratings. One
way, we discovered, was to cheat: play with 2 accounts, always letting 1
account win ... play 99 point 1 game matches ... never complete losing
matches, etc. Another "legal" way that came to mind was to consistently
play weaker opponents. If I understand the ratings formula, any 2 players
with a minimum of 400 experience points risk (winning/losing) an equal # of
points in any given match.
I made a consious decision to invite players with high experience, and
ratings equal to or greater than my own. Dice luck factor aside, I wanted
my ratings as accurately as possible to reflect my "true" level of play,
not any possible ego needs for (artificially) higher ratings.
I think that fibs is the greatest thing since sliced bread. I believe that
what marvin, Ran & company (sounds like a New York Law firm :) have
accomplished is so unique and appreciated, that asking any more of them
would be ungrateful. Having been a New Yorker most of my life, I tend to
be cynical at times ... always looking for a hidden catch ... no such
things as a free lunch sort of thing. Well, after 3 odd months on fibs
(some of them VERY odd), I'm happy to reprt that I'v been able to find no
hidden catches, other than the
fibs-is-down-oh-god-I-have-to-go-cold-turkey-again phenomenon. But I'm not
complaining :)
And at the risk of contradicting my above statement, I wholeheratedly
endorse Drago's suggestion!
>In rec.games.backgammon ra...@news.cs.columbia.edu (Dragomir R. Radev)
said:
Martin Ronek
Marty (NYCGuy - fibs)
Albert Steg
(Martin Ronek) wrote:
Another "legal" way that came to mind was to consistently
> play weaker opponents. If I understand the ratings formula, any 2 players
> with a minimum of 400 experience points risk (winning/losing) an equal # of
> points in any given match.
You are correct that two players with experience of 400+ each always
exchange an equal number of points, but you (and Dragomir) seem to have
missed the vital additional feature that if the underdog should win, *more
points* are exchanged than if the favorite should win.
If players rated, say, 200 pts apart play ten 3-pt (or whatever) matches,
and each player wins five matches, the underdog will come out having
gained ratings points, while the favorite's rating will have dropped by
the same amount. The ratings formula is drawing their ratings closer
together, in keeping with the fact that they seem to be performing equally
well.
> >In rec.games.backgammon ra...@news.cs.columbia.edu (Dragomir R. Radev)
> >...two people with the same rating, let's say 1610 rating might have had
> >completely different programs in terms of the strength of their opponents
> - one may have played only 1200-rated players, whereas the other may have
> >played only 1800-rated players.
> >An interesting statistic would be to have for each player, the average
> >rating of his opponents.
In light of what I wrote above, you should recognize that the average
strength of any two 1610 players' opponents is irrelevant! I don't have
the maths at hand to give hard numbers, but in order to maintain a 1610
rating against 1200-level opponents, a player must win a very high % of
matches ---much higher than necessary to maintain a 1620 rating against
more highly rated players. *If* ratings are considered a true reflection
of playing skill, then a profile of a players' opponent strength should be
irrelevant to interpreting his or her scores.
Albert
Albert
Dragomir R. Radev
Albert Steg <as...@tiac.net> wrote:
[some interesting stuff deleted]
>In article <403pi8$6...@pipe3.nyc.pipeline.com>, mro...@nyc.pipeline.com
>(Martin Ronek) wrote:
>
>Another "legal" way that came to mind was to consistently
>> play weaker opponents. If I understand the ratings formula, any 2 players
>> with a minimum of 400 experience points risk (winning/losing) an equal # of
>> points in any given match.
>
>You are correct that two players with experience of 400+ each always
>exchange an equal number of points, but you (and Dragomir) seem to have
>missed the vital additional feature that if the underdog should win, *more
>points* are exchanged than if the favorite should win.
>
Good point, Albert. It doesn't correctly describe my experience
though. Here is why: I have made a statistical analysis of the last 450 or
so points that I have played for. It seems that I have won many more rating
points against weak opponents than against strong opponents. Here are the
results:
against weaker players: 241-107
against stronger players: 51- 35
That is, I have won many more ratings points per point played against
weaker players than against stronger ones.
Note: in this message, "weaker" and "stronger" refer to ratings on FIBS.
--
michael rochman
>Different players play against opponents of different strength. So, two
>people with the same rating, let's say 1610 rating might have had
>completely different programs in terms of the strength of their opponents -
>one may have played only 1200-rated players, whereas the other may have
>played only 1800-rated players.
>An interesting statistic would be to have for each player, the average
>rating of his opponents.
Drago,
I second your idea. I've played a couple of guys that seem to have
knowledge and a sense of timing well beyond their points. And, I've
played at least one opponent with more points than me who I'd back
these other two in a money game against. Personally, my own points are
"soft" as I tend to try to stick with players around the same as me.
But, what happens is that I tend to get invited to play more often by
those with less points. I'd like to see your suggestion implemented,
if only to get a better picture of where I am.
Regards, Mike
STLguy on FIBS
michael rochman
>In light of what I wrote above, you should recognize that the average
>strength of any two 1610 players' opponents is irrelevant! I don't have
>the maths at hand to give hard numbers, but in order to maintain a 1610
>rating against 1200-level opponents, a player must win a very high % of
>matches ---much higher than necessary to maintain a 1620 rating against
>more highly rated players. *If* ratings are considered a true reflection
>of playing skill, then a profile of a players' opponent strength should be
>irrelevant to interpreting his or her scores.
Albert,
I'm not a rocket scientist, so I can't back this up with stats.
However, I know my matches have been weighted with opponents from a
little above me to a hundred plus, below me. In short, my rating at
this moment is 1802. But, I know I am not in the same league with
other 1800 + players in experience; Perhaps not with most 1700 +
players.
Using your premise, were I to constantly play matches with players of
about 1500 points, my point count would eventually diminish. I don't
buy into that. I think I can win perhaps 8 out of 10 matches against a
1500 point opponent, thus continuing to increase my point totals.
(I usually play a 7 point.)
However, were I to play 10 matches with a solid 1675-1725 caliber
player, I might win half of them. In that case my total would decrease
until such time as water sought it's own level.
Albert Steg
> I'm not a rocket scientist, so I can't back this up with stats.
> However, I know my matches have been weighted with opponents from a
> little above me to a hundred plus, below me. In short, my rating at
> this moment is 1802. But, I know I am not in the same league with
> other 1800 + players in experience; Perhaps not with most 1700 +
> players.
>
> Using your premise, were I to constantly play matches with players of
> about 1500 points, my point count would eventually diminish. I don't
> buy into that. I think I can win perhaps 8 out of 10 matches against a
> 1500 point opponent, thus continuing to increase my point totals.
> (I usually play a 7 point.)
No. I didn't mean to imply that.-- If you were to win a flat _50%_ of
your matches against a 1500 player, and played long enough, your ratings
would diminish, and his would increase until you and the opponent had the
same rating! If indeed you are able to win _80%_ of your matches against
a 1500 player, you may well increase your rating until you reach some
stable figure (I don't know what) where (80 x rating change when you win)
= (20 x rating change when you lose).
> However, were I to play 10 matches with a solid 1675-1725 caliber
> player, I might win half of them. In that case my total would decrease
> until such time as water sought it's own level.
Yes -- If you believe that you are over-rated, this is true. The question
underneath your two scenarios is whether a player who can only win 50% of
matches against players rated 1675-1725 can also be an 80% favorite
against players rated 1500. I have no idea -- perhaps someone more
conversant with the rating system and maths can answer that one.
My guess is that your estimations are impressionistic, and may well be
true to yoour experience, but may be colored by several factors: informal
nature of your data "sample" (anecdotal recall?); the fact that your
opponents are not always at their "true" rating; simple effect of lucky
streaks against opponents of any rating.
I stand by my original point, though, that the avaerage ratings of one's
*opponents* is largely irrelevant. The ratings formula, in taking into
account the ratings of the two players, is meant to level the playing
field by awarding a higher transfer of points when the underdog wins. It
should be just as hard to earn an 1800 rating playing against 1400-1600
players as it is to do so playing only 1600-1800 players.
Albert
James Eibisch
>Using your premise, were I to constantly play matches with players of
>about 1500 points, my point count would eventually diminish. I don't
>buy into that. I think I can win perhaps 8 out of 10 matches against a
>1500 point opponent, thus continuing to increase my point totals.
>(I usually play a 7 point.)
This seems a little optimistic! AIUI, a world-class player has something
like 70% chance against a competent/average player. Off the top of my
head, I would give a genuine 1700 player something like 58-60% against a
genuine 1500 player. These figures are almost certainly not completely
accurate, but I'd be surprised if they were way off.
--
_
James Eibisch ('v') N : E : T : A : D : E : L : I : C : A
Reading, U.K. (,_,) http://metro.turnpike.net/J/jeibisch/
=======
michael rochman
>My guess is that your estimations are impressionistic, and may well be
>true to yoour experience, but may be colored by several factors: informal
>nature of your data "sample" (anecdotal recall?); the fact that your
>opponents are not always at their "true" rating; simple effect of lucky
>streaks against opponents of any rating.
Albert,
Yes. Totally impressionistic. And at least just a tad colored as to
necessary modesty, too. :-)
However, I think we are missing something here. Would you not agree
that there is a far greater difference in skill level between a 1500
opponent and a 1600 opponent, then there is between a 1700 and an 1800
opponent? If so, then some adjustments might be made in the leveling
theory.
Mike
Albert Steg
> Albert,
>
> Yes. Totally impressionistic. And at least just a tad colored as to
> necessary modesty, too. :-)
>
> However, I think we are missing something here. Would you not agree
> that there is a far greater difference in skill level between a 1500
> opponent and a 1600 opponent, then there is between a 1700 and an 1800
> opponent? If so, then some adjustments might be made in the leveling
> theory.
Your point is that as players get better, a 100 point spread signifies a
smaller difference in strength? I confess I haven't a clue -- neither in
the realm of perceiving a difference in playing various people on fibs or
understanding whether the ratings system supports such an assertion.
An observation: When we look down from whatever pinnacle of BG knowledge
we have achieved, we see a vast gulf of understanding between ourselves
and our less experienced opponents. At the same time we look up to
stronger players and say "Eh. So that guy has maybe a small edge on me."
It's often hard to perceive how much we *don't* know.
I'm not really sure what point you are trying to make about the ratings
formula, though. All I was trying to impress on the readers was the
dynamic of differential points award depending on whether the favorite or
the underdog won any given match. Is your point that The gulf of
expertise between an 1800 & a 1500 player is worth *more* than the
differential payoff suggests? In that case it would be beneficial for the
1800 player to beat up on 1500-level players...but again, I don't know if
this is the case.
I do agree with the earlier poster who suggested that an 80% win
expectation for an 1800 player against a 1500 player seems optimistic.
What does thew FIBS formula offer as the favorite's edge in such a match?
Albert
michael rochman
>Your point is that as players get better, a 100 point spread signifies a
>smaller difference in strength? I confess I haven't a clue -- neither in
>the realm of perceiving a difference in playing various people on fibs or
>understanding whether the ratings system supports such an assertion.
Albert,
Off thread a bit for both of us, but yes, true.
>An observation: When we look down from whatever pinnacle of BG knowledge
>we have achieved, we see a vast gulf of understanding between ourselves
>and our less experienced opponents. At the same time we look up to
>stronger players and say "Eh. So that guy has maybe a small edge on me."
>It's often hard to perceive how much we *don't* know.
But, of course. Happens in any game that contains both skill and luck.
>I'm not really sure what point you are trying to make about the ratings
>formula, though. All I was trying to impress on the readers was the
>dynamic of differential points award depending on whether the favorite or
>the underdog won any given match. Is your point that The gulf of
>expertise between an 1800 & a 1500 player is worth *more* than the
>differential payoff suggests? In that case it would be beneficial for the
>1800 player to beat up on 1500-level players...but again, I don't know if
>this is the case.
Nope, my point is that a 1500 player on FIBS may be either a rank
beginner or very close to one. I don't think a rank beginner can beat
a 1700+ player more than 2-10 matches to 7, which is what I most
always play. Not games....matches.
James Eibisch
>James,
>1500 as rated by whom and where?..is the question. I don't know that
>I'm a "genuine 1700" player. Whether I'm better or whether I'm worse.
>I stick by what I said.
>Here's why...
>Let's suppose that we were setting up a mythical backgammon server.
>Let's also suppose that we were to invite 500 of whom are generally
>considered to be the finest players in the world.
>Let's let them play for a while. The cream will rise to the top.
>However, the bottom group..the 1500ers will be far superior to the
>1500ers on FIBS, where many of whom are beginners.
>Thus my suggestion that 8-10 by a 1700er against a 1500er on FIBS is
>not altogether unrealistic.
Yes, but a beginner to bg is not a 1500 player - he will stay at around
1500 only for a very few games and will soon drop much lower. I don't
know how strong a player would have to be before he had 80% chances
against a 1500 player, not counting those with very low experience
(perhaps this should have been in the equation from the start), but it
would be a very strong player indeed.
The ratings system is almost like an organic entity - it's self
contained and uses negative feedback to contain itself in some order
around an floating median. Probably. The ratings aren't absolute,
they're relative, so they are meaningless when compared with the ratings
of another, mythical FIBS. FIBS ratings cannot be used in any context
other than the FIBS ratings system as used by the current active
population of fraggel65.mdstud.chalmers.se 4321.
I haven't worked this out with the ratings formula yet, but until then I
believe that a player who maintains a reasonably-true 1700 rating has
nowhere near 80% chances against a player who maintains a
reasonably-true 1500 rating. I'm rated around 1500; maybe we could play
10 games :)
William C. Bitting
: mroc...@mo.net (michael rochman) wrote:
: >James,
: >1500 as rated by whom and where?..is the question. I don't know that
: >I'm a "genuine 1700" player. Whether I'm better or whether I'm worse.
: >I stick by what I said.
[SNIP]
: Yes, but a beginner to bg is not a 1500 player - he will stay at around
: 1500 only for a very few games and will soon drop much lower.
[SNIP]
Clearly the EXPERIENCE NUMBER is key to judging the validity of a rating
number. While all FIBS players start at 1500, because of a multipler to
the rating changes, which starts at 5(!) and declines to 4 at an
experience of 100; 3 at 200, etc., 1500 is not what many FIBSsters will
have as their rating when their experience is 200, not to mention 2000.
Some players decline to near 1300, and lower, quickly. Others like
'loner', 'mloner, 'STLguy', etc., quickly pass 1700 or 1800, not to
mention 1900 and 2000 by the former two!
So, while it is true that a 1700 or 1800 rated player may win 8 of 10
seven point matches against many new FIBS players, that won't be true
when the 1500 rated, 0 experience, player is one of the loner's, etc..
Nor will it be true when the 1500 rated player has a couple of thousand
experience points unless, of course, the player 1700 to 1800 player is
somebody on the way to 1900 or 2000.
The following table is the "batting average" the favorite must achieve to
maintain his rating at various rating differences. For players whose
experience level is over 2000, it may have considerably more validity
than for players whose experiece level is under 400. Players with
experience levels under 500 and ratings of mid 1700, not to mention over
1800, may be still be underrated simply because of not having played
enough matches!
^L
BATTING AVERAGE FOR FAVORITE TO MAINTAIN HIS RATING.
(For the underdog, his average is 1 minus the favorite's.)
rate points in match
diff 1 3 5 7 9 11
---------------------------------------
0 .500 .500 .500 .500 .500 .500
100 .529 .550 .564 .576 .585 .594
200 .557 .598 .626 .648 .666 .682
300 .585 .645 .684 .714 .738 .759
400 .613 .689 .737 .772 .799 .822
500 .640 .730 .784 .821 .849 .871
600 .666 .768 .824 .861 .888 .908
700 .691 .802 .858 .894 .918 .935
If the proposition is that a player is going to win 8 of 10 seven point
matches against EXPERIENCED 1500 rated players, that batting average is
.800. From the above table we might therefore concluded the rating of such
an .800 hitter should be about 450 points over 1500, or 1950! Thus a
player who correctly guages that his batting average is .800 against
experienced 1500 rated players, may just as well judge that he is
underrated until his rating reaches 1950 rather than that the 1500
player is overrated.
If the propostion was winning 10 while loosing 2, then the batting
average is .833, and the batter should be above 2000 on his rating index
given enough experience!
Good Dice! wcb on FIBS
michael rochman
>I haven't worked this out with the ratings formula yet, but until then I
>believe that a player who maintains a reasonably-true 1700 rating has
>nowhere near 80% chances against a player who maintains a
>reasonably-true 1500 rating. I'm rated around 1500; maybe we could play
>10 games :)
James,
Matches to 7 or more, not games. :-)
Mike
michael rochman
>Some players decline to near 1300, and lower, quickly. Others like
>'loner', 'mloner, 'STLguy', etc., quickly pass 1700 or 1800, not to
>mention 1900 and 2000 by the former two!
Hiya Bill...
First of all...I should be so lucky as to be included in the same
paragraph with those others. :-)
>So, while it is true that a 1700 or 1800 rated player may win 8 of 10
>seven point matches against many new FIBS players, that won't be true
>when the 1500 rated, 0 experience, player is one of the loner's, etc..
>Nor will it be true when the 1500 rated player has a couple of thousand
>experience points unless, of course, the player 1700 to 1800 player is
>somebody on the way to 1900 or 2000.
Ok, after peeking ahead at your chart (thanks for posting it) I must
say that my gut instinct must be incorrect.
>BATTING AVERAGE FOR FAVORITE TO MAINTAIN HIS RATING.
>(For the underdog, his average is 1 minus the favorite's.)
>rate points in match
>diff 1 3 5 7 9 11
>---------------------------------------
> 0 .500 .500 .500 .500 .500 .500
>100 .529 .550 .564 .576 .585 .594
>200 .557 .598 .626 .648 .666 .682
>300 .585 .645 .684 .714 .738 .759
>400 .613 .689 .737 .772 .799 .822
>500 .640 .730 .784 .821 .849 .871
>600 .666 .768 .824 .861 .888 .908
>700 .691 .802 .858 .894 .918 .935
BTW, Bill, I can't even beat you, yet. :-)))
And, we have a match to finish and you are ahead.
James Eibisch
>Matches to 7 or more, not games. :-)
Yes, I think I missed your point about 7pt matches before, and was going
on money games. I'm not sure I fancy losing that many matches ;)
michael rochman
>mroc...@mo.net (michael rochman) wrote:
>>Matches to 7 or more, not games. :-)
>Yes, I think I missed your point about 7pt matches before, and was going
>on money games. I'm not sure I fancy losing that many matches ;)
James,
But you won't. :-)
Mike