Skip to content
Little GolemMake your next move.

Updated FAQ

Public forum discussion.

General

#51 · Art Duval

9 May 2007

I'm pretty sure ratings are computed based on ratings as the game finishes. Can someone confirm? Probably the easiest way to check is what happens to your rating as you win against someone who is in the process of losing a lot of games due to timing out.

#52 · Dvd Avins

9 May 2007

Yes, I know from experience that the ratings are calculated using the old ratings just before the game is finished. That's also the way it was always done in postal chess, and it's what maes sense for our current version of correspondence play.

#53 · ypercube

9 May 2007

Yes, about 2, ratings cahange is based on ratings when the game finishes.

#54 · isometry

9 May 2007

I confirm everything this is being said about ratings. :-)

#55 · ypercube

9 May 2007

About the rating formula and StreetSoccer, the "result" is YourPoints/maxPoints, where maxPoints is 5, e.g.:

1 when you win 5:0

0.8 when you win 4:1

0.6 when you win 3:2

0.4 when it's 2:2 or you lose 2:3

0.2 when you lose 1:4 and

0 when you lose 0:5

#56 · Hjallti

9 May 2007

I didn't check but if you said so in the faq but K is less than 32 if your opponent has nonge or only few games played. Richard never gave a clue on what 'less' or 'few' in the previous sentence means. The new player gets K=32 (playing against a player with enough games, that is)

((I think EinStein was create for best of 50 or so matches certainly not for single games. I think Ingo mentionned this in the EinStein forum somewhere))

#57 · isometry

9 May 2007

Hjallti, I think the system has changed and that it no longer matters whether the players have few or many games. My new nick apparently took rating points from established opponents already in the first games.

#58 · Hjallti

10 May 2007

Do you mean they lost a lot of points? In that case you might be right

The rating graph of your nick can't tell you anything about this... only the rating change of the other players. (If it the difference is not the same in both changes their K's are different (perhaps 1 difference could come from rounding up or down))

#59 · isometry

10 May 2007

Yes, they lost points; I believe exactly the same as I gained (or the other way around when I lost).

#60 · Greck

11 May 2007

it deserves first page place :)

#61 · Ed Collins

12 May 2007

Thanks.

Okay, I've again updated the rating section, this time with ypercube's info about StreetSoccer.

I think I've got it straight but I still might have something wrong.

#62 · isometry

12 May 2007

So with 2:2 it could happen that both players drop in rating? Is it not counted as a draw (0.5)?

Under the rating section there is a small paranthesis error...It says: "(Note: ..." but there is no ")".

An even smaller error: The list of games on top of the page should read TwixtPP, not Twixt.

#64 · Marius Halsor

12 May 2007

I think indeed with 2:2 both players could drop in rating. Haven't tested it, though...

#65 · Greck

12 May 2007

it happened to me once. The tip is that with equal rating the expected is 2.5 of 5 If you have less than expected, you lose points :)

#66 · Hjallti

16 May 2007

I keep on saying great work!

Putting this back up front

#67 · wccanard

24 May 2007

I have a comment about the formula for the rating. The formula is a standard one, used one other servers as well as this one, but there are two times in it where something is raised to a power, and it's very difficult to see this in what is written in the FAQ (at least on my browser!). Has the formula been incorrectly typed in in the FAQ or is my browser not showing up one of these "raise-to-the=power"s? My reading of the html source is that the author of the FAQ has forgotten one of them. Let me get more precise (and probably more dull, for most of the people reading).

The first time in the formula when you raise something to a power is when you're trying to construct a number which is bigger than 1, and furthermore a *lot* bigger than 1 if the rating of player 2 is much higher than the rating of player 1. The way this is done is

1 + 10 ^ ( (r2 - r1) / 400 )

that is, 1 plus 10 to the power of (something or other). Of course I'm writing r2 for Rating2 and r1 for Rating1. In the FAQ, at least with the settings my browser has, this raising to the power is indicated by a *very small* rise in the text: r2 - r1 / 400 is a tiny bit higher than the 10. Not ideal, but at least not wrong.

But there is another place where you raise something to the power: that number we just constructed above, we don't really want something like that: we want something a bit *smaller* than 1, and a lot smaller than 1 if r2 is much bigger than r1, so we take the reciprocal. That is, we take that term I just talked about above, put it in square brackets [...], and then raise it all to the power -1. I am looking at the FAQ now and I can't see this second power indicated at all. The -1 (that last -1, right on the right of the formula) should be up in the air above that bracket. Is this a font issue at my end or has the author of the FAQ forgotten to put that -1 up in the air?

Let me try and explain the logic of the formula. The number

1 + 10 ^ ((r2 - r1) / 400)

is exactly 2 if the two players have the same rating, is bigger than 2 if player 2 has a higher rating than player 1, and is between 1 and 2 if player 2 has a lower rating than player 1. You then work this out, take the reciprocal, and get a number which is between 0 and 1/2 if player 2 is rated higher, exactly 1/2 if the two players have the same rating, and between 1/2 and 1 if player 1 is rated higher. You then do w minus this, and if you think about all the cases you're now getting a number which is positive if player 1 has done well (e.g. he has won, or drawn against a player that is better than him) and negative if player 1 has done badly (e.g. he has lost, or drawn against someone with a worse rating). So this number reflects how his rating should change. But it's between -1 and +1 which is too small, so multiply it by 32 to scale it and then add it to the old rating to get the new. Simple!

Another note: why introduce the variable K if on this site it is always 32? Why not just put 32?

wc

#68 · wccanard

24 May 2007

Ooh---sorry---by "w" in the para above I meant "Result". I was staring at the formula on Richard's PBEM server, where it is correctly typeset (in the "ratings" help file, for anyone who uses that server).

#69 · ypercube

24 May 2007

K was intoduced because there was suspicion that different K was used for some time in one of the games.

#70 · wccanard

24 May 2007

In Richard's PBEM server different (smaller) K's are used for players who haven't been around for too long, because their rating might not accurately reflect their ability. I have no idea what the system is here (I haven't been around for too long :-) ).

#71 · ypercube

24 May 2007

You are right about the -1.

It should be -1 .

#72 · wccanard

24 May 2007

ypercube: right. I just didn't know how to embed the html like you did :-)

In practice I use a "rule of thumb" to calculate ratings: if I beat someone I'll get 16 points, plus about (difference in ratings/ 25) extra points, where by "difference in ratings" I mean "his rating - my rating", so e.g. if I beat someone rated about 200 less than me I should expect to score about 16-8=8 points. This "simplified" formula is not too hard to calculate in your head and gives quite accurate answers unless the difference is large.

#73 · isometry

24 May 2007

There used to be a smaller value for an established player playing against a new player, but apparently the value is now 32 regardless of how many games the players have finished.

#74 · isometry

24 May 2007

If you want to write a table of the intervals for the rating differences that give a change of 16, 15/17, 14/18 etc. points, just calculate

400 log_10 (32/r - 1)

for r = 15.5, 14.5, 13.5 etc.

For example, r = 15.5 yields 10.86 and r = 14.5 yields 32.67. This suggests that a rating difference in [0, 10] gives a change of 16 points, while a rating difference in [11, 32] gives a change of 15 or 17 points (depending on whether the higher of lower rated player wins).

#75 · Dvd Avins

24 May 2007

Further explanation:

The "two times" wccanard metnions are only one, as far as I can see. But the number needs to be lower than 1 in some instances and greater than 1 in others. (And exactly 1 if the ratings are equal.)

The (r2 - r1) / 400 part of the formula is the odds, let's call it n, where the odds of winning a game in traditional gambling parlance is n to 1.

The K (32 here and in most places where such systems are used) and the 400 are really 1 parameter, not 2. Someone less distracted than I am at the moment can do the algebra to isolate the actual paramenter, where one of K and 400 is raised to the power of the other. But that combination of the two is isolatable, and determines the spread of the rating system. The greater the spread, the less a difference of 1 rating point means.

#76 · Dvd Avins

24 May 2007

Iw was wrong. It's the 10 and the 400 that are a single parameter. The K is seperate.

#77 · wccanard

24 May 2007

isometry---are you sure that the rounding rules are really what you implicitly suggest? It's a natural guess but I don't think it's what is implemented. Try some experiments on some data you have and see if you agree with me.

As an example, I beat someone at dots and boxes who was ranked 256 points below me, and my rating increased by 5 points. However the precise formula gives 5.96 or so for the change in rating, so my guess based on that one stat is that the rounding is always going down and you should try r=15,14,... rather than what you suggest above. Do you agree?

Anyway, my earlier point is that in the region r=[6,26], the graph of the function you write above is close to being linear with slope -25; that's why my rule of thumb works.

wc

#78 · wccanard

24 May 2007

Dvd: you are right about the 10 and the 400: one could rephrase it as

c ^ (r2-r1)

where c is some constant that is close to, but a little bigger than, 1. The 32 really is a different parameter, as isometry's inverse formula also shows. Because the same number is apparently being used for all people, it just is some scaling factor: e.g. if it had been set to 16 at the beginning of time rather than 32, then everyone's rating would be pretty much exactly half the distance away from 1500 than it actually is (modulo rounding errors). The 10 and the 400 are more subtle constants; if you tinkered with them then in practice it would mean that the relative importance of beating players ranked much higher than you would change. I'm sure you know all this already. I think that the choices made on this server have stood the test of time in the sense that they are used in several other places too and seem to give reasonable results in most cases [cue a huge stream of people saying "they didn't give reasonable results on this occasion (rant)!]

#79 · wccanard

24 May 2007

PS the formula as it is typeset on ypercube's player page looks just right to me, so the author of the FAQ might want to look there for a simple explanation of what I am trying to say---that -1 at the end should be smaller and up in the air :-)

ypercube: if you're reading, why don't you forget about this basic algebra and answer my questions on the dots and boxes forum :-)

#80 · isometry

24 May 2007

You may be right about the rounding.

#81 · Dvd Avins

24 May 2007

I wouldn't call the 32 simply a scaling factor. It's relation to your c (which is 10 ^ .0025 or 1.00577306 as implemented) determines how much recent results matter, as opposed to older results.

c is a scaling facor. The larger it is, the more a single rating point means.

#82 · Hjallti

5 Jun 2007

Somebody (else then Richard) knows when these will be up.

By the way... I don't see how 32 would make any differnece to, for instance, 16 or 128 in how much older results matter. It would only make the scope of results bigger or smaller. The influence is the same, since it is pretty linear to go from a scoring based on 32 to one on 128... (even regardless of the starting score of the players being 0, 1500 or 1000000,...)

#83 · Dvd Avins

5 Jun 2007

10 ^ (1 / 400) determines the worth of a rating point -- the spread of the distribution. Currently, players of equal rank each bet 16 points on a game and the winner takes 32. In all cases, the total of the bet is 32, but the higher rated player (if there is one) antes up the greater portion of the pot.

If you increase the size of game pots, you increase the volatility of the ratings, but no the value of a rating point.

#84 · Hjallti

6 Jun 2007

so?

Of course the word 'in(de)crease' can mean two things, my comment is about comparing to rating systems which stay the same but differ on the 32.

If you would change today the 32 into 64, on this website, of course the influence would change, My point that it doesn't make a difference if you initially would have started with 64 rather than 32. (except for the last digit for rounding the numbers eventually somewhere).

Let's say everyone started at 1500 with 64 as number compared to 32

then a player would have 2000 if has 1750 now. It is just the double (difference from the start), it does not matter if the rating graph would show a fast climb in the beginning and merely status quo now, or the a slow climb.

Your assessment implies that the order of the rating would be different (In Go and chess where people are asked to give their initial rating of course it would matter a lot.)

#85 · Hjallti

6 Jun 2007

With order of rating I mean: Can you give an example that two players would end up in a different order (one more another less) with everything equal except the 32?

#86 · isometry

6 Jun 2007

_Let's say everyone started at 1500 with 64 as number compared to 32

then a player would have 2000 if has 1750 now._

That is not correct. In order for everything to double like that, the 32 would have to be replaced by 64 - true; but we would also have to replace the 400 by 800.

If 32 is replaced by 64 but everything else is left untouched, then old matches matter less and recent matches matter more. Think of the opposite scenario: 32 is replaced by a very small number, say 2. Then a series of several good results will not alter your rating a lot - the old matches (and rating) will still have a great influence.

The 400 is the expected rating difference between two players if one of them should beat the other in 1 out of 11 games. It doesn't matter if the 32 is replaced by a different value.

#87 · Hjallti

7 Jun 2007

Got the point,

sorry for confusing everyone... you are right... I intuitivily thought of the formula, and forgot that in the technical formula that 400 was dependent (in other words I tought of 400 as P(1/11) or so, which is dependent of 32, and would become 800 automatically), Stupid mistake of mine

#88 · David J Bush

7 Jun 2007

So how do you all like the spiffy new table Ed added for estimating rating changes?

#90 · wccanard

7 Jun 2007

I said the 32 was "a scaling factor". Dvd disagreed and, after isometry's explanation, I finally have understood his point! You're right Dvd, it's more subtle than a scaling factor, and sorry (esp to Hjallti) for the confusion I caused.

It's nice to know that Ed is listening, even if he's not posting. Thanks for fixing the typesetting of the formula---and nice table!

wc

#91 · Ed Collins

7 Jun 2007

For the record, David Bush supplied the information for the table for me.

Just about everyone has now a hand, in some way, in helping put the new FAQ together. It no longer can be considered MY FAQ. It is OURS.

:)

#92 · Ed Collins

28 Jun 2007

I'm bumping this, since it was almost moved back to Page 3.

#94 · alain

30 Aug 2007

My one comment is that the format of who invented it should be consistent. I like the way you've done it for the games at the top by having a small separate paragraph saying: Invented by: Inventors Name, Date.

Very nicely done! The LG FAQ should include a link to this excellent document.

#95 · Jonny

30 Aug 2007

Well, some things have to be added... Example: Championship: " If a new player sign for a championship he will be beginning in a level that is conform with his score. If he is in the TOP 5 he will come in the first level." Something like this... but just better and corecter in speaking ;-))

#96 · slaapgraag

21 Sep 2007

Just read the page - now i finaly understand 'SON'!

#97 · Ed Collins

22 Sep 2007

Thanks. I'm glad that example I recently added helped somebody.

#99 · David J Bush

19 Oct 2007

Someone just asked me a question that wasn't in your FAQ!

"What do all those numbers describing a tournament mean. (ie. hexch3.2.1)?"

And I have one:

"When is Richard Malaschitz going to provide a link from LG to your page?" :-)

Last page