Thad: That is my guess as well. It is much more probable that the error lies in wrongly reusing already used dice than in the random number generator.

grenv: Even easier and more reliable, since all games all recorded : querying the database for all pair of first roll and second roll in every game since x months. I'd expect the results to be overwhelmingly abnormal.

Pedro Martínez: LOL, I decided not to play any more tournament but that one had me really tempted !

What about a Logik tournament with compulsory random placement (see bug 1487) ? Brainking, the site with Easter-egg game variants.

21 : 13/11 6/5
31 : 8/5 6/5
41 : 13/9 8/7
51 : 13/8 6/5
61 : 13/7 8/7
32 : 13/11 8/5
42 : 8/4 6/4
52 : 13/8 6/4
62 : 13/5
43 : 13/10 13/9
53 : 13/5
63 : 8/5 8/2
54 : 13/4
64 : 24/14
65 : 24/13

My propositions. I didn't check with a bot and I could be wrong about some.This of course would be with 100% of same rolls. Some of them are too bold to be played against only 50% of same rolls.

TC: I noticed the bias too. Since it is quite hard for us, but very easy for Fencer to generate large statistics, I think it should be investigated asap. It is a potential exploit. For instance it makes 13/5 with an opening 62 much better than the usual 24/18 13/11 (until now I refrained from using that exploit).

My guess was that it had to do with the player at move winning or losing the opening roll in a multi-game match. But I didn't check that.

On a remotely related note, very long ago I submitted bug #1487 about a bias in the random placement in Logic, based on a statistic of 50x5 peg generations. I didn't even get an answer. I think that Fencer sometimes commits the fallacy of generalizing "some of my users are idiots" to "all my users are idiots".

spirit_66: Analyzing your played games with GNU is indeed an efficient way to catch the cheaters, but since it is not that rare even for casual players to play one BG game close to perfection you need to be a little more quantitative :

>BG games are analysed by GNU

How many games ?

>and one can see that ones opponent didn't make any mistake

Didn't make any mistake greater than what threshold ? GNU doesn't color mistakes lower than a certain threshold.

> and what if this happens again and again???

How many times ?

E.g. I would certainly take an aggregate error rate below 1 (Snowie) or 2 (GNU) based on two or three consecutive seven-pointers as an evidence for bot cheating.

paully: That's odd, you make it sound a bit like you are bragging about driving while drunk

wetware: Since I counted only two double 5s (the usual one and the former 22) instead of four, I probably missed twice 10 pips, so the grand total would be 336 + 20 = 356, and the average 9 8/9. Better ?

wetware: You are right ! I spotted the trap with one die and fell into it with two dice :-)

wetware: With both dies changed : [ 294 ( = 36 * 7 + 6 * 7, the usual sum of all rolls) + 30 (10 rolls with a 2 changed to 5) + 12 (1 roll with a double 2 changed to a double 5) ] / 36 = 336/36 = 9 1/3. Enough to win most of the long races !

I will leave the "one die" to someone else, but beware, it is higher than halfways between 8 1/6 and 9 1/3.

alanback: Exactly. One needs to be sure that one cannot get advantage by resigning or timing out, not that one cannot be disadvantaged by it. The eventually of being forced to play one or two more moves is not much in comparison of being forced to play all those forced moves which could be auto-played...

alanback: It is the only one as far as I can see. I can't think of any position where there is contact and a backgammon cannot possibly be lost.

But thinking again about single games vs gammons I now see that "one checker off" is not a necessary criteria for gammons to be impossible. If all your checkers are one away from bearing off, and the opponent has more than 5 checkers on the board, one will manage to bear at least two checkers off whatever happens.

So awarding a gammon unless there is one checker off is not 100% conceptually clean after all. It is still probably the best see-in-one-glance estimate. Who would like a sophisticated software to decide about the result of a game ?

playBunny: The parallel is that in chess like in backgammon, there are more than 2 possible results (win or loss). In chess there is the draw as well. So there is the question of what result you get after you time out. And the universally accepted solution in chess is that you get the worse possible result - that is, you get a draw if and only if the rules make impossible for you to lose.

Porting that principle to backgammon would mean that :
- By default, timing out costs a backgammon.
- If a backgammon cannot possibly be lost from the position, timing out costs a gammon.
- If a gammon cannot possibly be lost from the position, timing out costs a single game.

playBunny: I agree with everything you said, except that awarding a backgammon when it is still possible seems conceptually clean to me (if not ideal, nor necessarily fair). In chess, if your opponent loses on time and you have one single pawn against an army, you still win the game (if you don't have the pawn, it's a draw).

But the way it is seems acceptable to me, and I certainly wouldn't make a bug record or feature request of it, since there are more important pending ones about backgammon.

playBunny: Generously awarded gammons ?

But I checked it (easy, just try to resign a game with contact and with no checkers in the opponent's home board, then don't confirm the resign), and you were right, only a gammon is awarded in that case. Somehow my memory played a trick on me, sorry for confusing the debate.

Now it is also true that it seems near to impossible to think of a position with no checkers in the opp's board, where one would be better off resigning a gammon. But IMHO the possibility to do so is still a conceptual bug. And indeed, here is a somewhat contrived situation where it can turn into an exploit :

Triple gammon tournament, there is only one game left to play between the leader and the second. The leader is 4 points ahead of the second, so that he wins the tournament unless he loses a backgammon. To avoid that, he manages to leave the opponent's home board, resigns and wins the tournament. That's wrong.

playBunny: If you resign when you're clear of your opponent's home then you'll be asked to confirm losing a gammon.

Yes, except that you need to be clear of your opponent's home AND there needs to be no contact any more (otherwise a backgammon could still be lost after a hit). This is how it should be and to the best of my memory how it is - I will soon have the occasion to check it for real :-(

Gordon Shumway:
It might be easy to integrate it, but it would be very bad style as it would mean that BrainKing first produces a broken mat code and then reads and manipulates it to remove the bugs, instead of not producing them in the first place. And I don't even think fixing the bugs right where they are would be any harder

Shshshsh ! This is what I was trying to subliminally suggest

A lot of numbers are the same in every line, so I guess that this is just the routine position setup. the last two digits are the dice rolled. But all that stuff in the beginning?

Yes, it is a routine position setup. I didn't know about the Snowie .txt format though, and thought that inserting the same string everytime would do it. The dice rolled shouldn't be of importance since they are given with the next move anyway. Having the good values in the first places is probably more important than I thought though.

Gordon Shumway: Two more thoughts :

- As it is written in Java, it should be easy for Fencer to integrate it in his .mat generation code.

- Errr... do you have any plans to fix the Nackgammon files as well ?

Gordon Shumway: That is GREAT ! Your tool seems to work perfectly and is very easy to use. Especially helpful is the possibility to fix many files in one batch.

Thank you so much, this gives me some motivation to play more backgammon on Brainking again !

playBunny: "Top in ToyTown", I like that expression !

alanback: IMHO there are exactly three points that should be fixed in order to make a great backgammon site out of BK, where the experts would love to play :

1) Autopass and autoplay, regardless of the settings of the opponent, and which can trigger one way of the other even when cubing is an option.
2) A fair rating system.
3) A working .mat export.

2) and 3) are really easy to implement. The .mat export is almost done but still has an annoying bug, as for the rating system, adapting the BK rating system is just as easy :

A) Before putting the rating difference into the BK formula, multiply it by the square root of the match length (and then divide it by 5 if copying the FIBS formula, but that is not essential).
B) After calculating the gain or loss in rating, multiply it by the square root of the match length (and divide it by 5 if wished).

coan, you are right that Fencer might recalculate all ratings if he changes the rating formula, which needed a downtime the last time he did it. If he does it, I bet that I will lose my #1 spot to XXXXXXX (the unnamed computer user).
But frankly, I don't really care whether he does it or not. One way or the other, once the system is corrected, the ratings will converge to their accurate values in the long term, and what I care about is the long term.

Andersp, you are wrong in thinking that the problem lies within the higher-ranked players playing between themselves. Actually, the surest way to exploit the current rating system is to play 21-pointers against low rated opponents, when one wins non negligible rating points with almost no risk.

alanback: When I talked to Fencer about that, he did not refuse to implement it, he sounded something like it would probably be a good thing, but of low priority. The priority might be higher if more people here would know that the present rating system is flawed for multi-point matches. So I am shouting it again : the present rating system is flawed for multi-point matches !

Hrqls: When bearing in it is best to aim for the points 4, 5 and 6. The low points can very well remain empty. The reason is that all dices below 4 will play fine even if the low points are empty (e.g. with a 3 you play 6/3, keeping a fine distribution). By contrast, an empty 4-point (for instance) is a nuisance, because all fours have to be played 6/2 or 5/1, accumulating checkers on the low points, that are likely to waste pips later.

So in your example, the move is definitely 8/6 8/4.

Hrqls: I think you are right, but that there is a difference only if you roll 2-1 or 1-1 with your next roll. If you roll 2-1, 6/4/off will likely win a roll in comparison to 6/1 5/2. If you roll 1-1, 6/4/off will likely lose a roll. Since 2-1 happens twice as often as 1-1 and you are way ahead (so after 1-1 you are very likely to win anyway), 6/4/off seems clearly better.
It is probable that if your opponent was ahead and you desperately needed doubles, 6/1 5/2 would be the correct move. I doubt there is much equity at stake in any case.

playBunny: You are right that a bot should also play unrated. I didn't think of mentioning it because of the BK rating system being *ahem* as soon as different match lengthes are possible.

Puckish: Frankly, I am rather touchy (who said paranoid?) about computer use, and I find it highly unlikely that computer help is widespread here in backgammon games. As I said, I rather think that the case I disclosed is an unique one. It is certainly a very different story on sites where money can be won.

The behaviour you described would also fit me. Like alanback, when facing a tough move, I often like to get a night of sleep over it, and my ideas are sometimes clearer the next day. Or I see that the move requires a lot of counting, and keep it for when I have the time and availability to conduct the whole calculation.

Moreover, the way GnuBg works, it is a bit tedious to enter a specific position, while it is very easy to open saved games and enter moves one after the other. So a cheater is more likely to consult the bot for every move, even though a "clever" cheater would make himself tough to detect by introducing enough mistakes to stay with a "human" error rate.

I see that my made-anonymous accusation is likely to open speculation. So I will at least say that I have had a PM conversation with the cheater about it quite some time before posting here, so he/she know about it, and nobody else should feel that he/she could be targetted.

playBunny: That is a consistent point of view, too

But I would add two further conditions in order to be a proper bot :
- It should be advertized as such in a very clear manner (even more accessible than the player profile).
- It should not enter tournaments, because no one should be forced to play against it.

I'd resign because I am already playing more than enough live games against GNU, I don't need to log on Brainking for that. But if you would still enjoy playing, that is OK with me too.

As for that person's ego, frankly, I don't know how it works. It is true that I would be more ashamed to achieve a high rating by being GNU's proxy than by having everybody resigning to me.

Binabik: I agree, except that nobody should resign against him/her before the information has been double-checked. I am sure of what I am saying, but I don't expect to be trusted automatically.

rod03801: The PM to Fencer has indeed been my first action. He didn't have the time to investigate the case by himself, and we all know how fond he is of backgammon. So he agreed with my proposal of seeking for confirmation by other players on the backgammon board.

Quite possibly I misunderstood him for naming the accused player, but I hadn't thought of any other way. Remember that ********** is mainly playing private games and that only his/her opponents are able to see them. OK, now maybe there is another way. Add the instruction 0) : PMing me to know the handle of the player.

coan.net: I apologize, I realize that I should at least have given a link to the player's profile instead of naming him or her. This way, everyone could have chosen to click it or not. But I guess it would not even have been OK this way.

The question is what we do about it. Nobody cares ? I shouldn't even have told ?

By the way, this is definitely not a general talk about cheating. It is a very specific one about one isolated case. On a general and more positive note, I would say that I am pretty convinced that all the other people I played backgammon with here were playing by themselves.

alanback: He does ! Never trust the BK rating :-)

Unfortunately I found out that ********** was using a bot to play his moves in Backgammon and Nackgammon games. As Fencer understandably won't to act on it until it has been confirmed by other users, and as ********** has mainly played private games, I appeal to his recent opponents that are acquainted with the usage of GNU backgammon (the best free program and the one that ********** uses). Here are the steps to reproduce :

1) Choose a recent finished private match that you played with ********** .

2) Click "download MAT code" at the top right of the game screen and save the file to your disk.

3) Start GNU backgammon.

4) Go to Settings / Analysis and choose the "Expert" level for both checker play and cube decisions.

5) Click the Import button, choose the .mat format and your saved .mat file.

6) Go to Analyse / Analyse match. GNU will analyse all played moves.

7) Go to Analyze / Game statistics - and not match statistics, because there is a bug in the BK import format that swaps players every two games. Browse the games. In each of them, I bet that you will see one player with a human error rate (you) and one with an error rate of virtually nil ( ********** ).

Before jumping to the logical conclusion, you should know that the world's best players have an "Equivalent Snowie error rate" around -2 on a good day. You can check that for yourself by downloading world-class matches from here : http://www.hardyhuebener.de/engl/matches.html

On a last note, "expert" is a lowish standard in GNU backgammon, and I have seen ********** play a move that was also played by the expert level, but discarded as a big blunder by the "world class" level (which analyses 2-ply instead of 0-ply).

Thanks by advance for your input !

********** Edited out users name. **********

mangue: The rules are that if you can play both dice, then you must do so. If you play the 5 first, then you are blocked and cannot play the 3 any more (while if you play the 3 first, you can bear off with the 5).

coan.net: Everything you said !

Plus making forced moves automatically, even when they are not "passes". That for all games without exception. This last request should probably give rise to a separate personal setting, since I guess that some Ludo players could be happy with autopass but not with autoplay.

grenv: It is complicated enough for me not to have dived into one method yet. The pips to add are not linear in function of the number of checker on either the 1 or the 2 points, gaps have to be accounted as well.

If you are interested, give a look to http://gammonline.com/members/Dec00/articles/pipples.htm . A bit too advanced stuff for me I must say.

Puckish: A very simple rule works fine in most situations. If on roll you are 10% (of your pip count) ahead minus 2 pips, you have a borderline double. If you are 10% ahead plus 2 pips, your opponent has a borderline take/pass.

That is of course in "money play". The take point can be different in function of the match score (and the presumed strength difference between the opponents) ; and when one of the opponents has too many checkers on low points.

wetware: Probably not deep enough :-)

DailyGammon's .MAT format is fine and works also with Nackgammon matches.

If you are willing to write such a program, what about doing it in Java so that Fencer can integrate it to BK ?

wetware: You are right, this is BK bug #1181, that I posted back in March :-(

Thad: I am not experimented enough to say it for sure, but I think that the recognized backgammon etiquette is to carry the game on as long as there is a theoretical chance of winning (even if it is a one to a billion shot), and resign as soon as there is no more. Usually one assumes proper play by the opponent when computing theoretical chances.

But as it is often longer to compute whether a theoretical chance exists than to play on, in live play one often simply carries the game to the end. In turn-based play, I think it is nicer to bother with the simple computing and resign on time. Note that the Dailygammon server implements that etiquette by resigning the game automatically when a win is impossible.

In matches one should normally be allowed to resign the type of game one decides to (normal game, gammon, backgammon) and the opponent should be allowed to reject one's resign offer if he finds in not high enough. This is not possible on BK though.

Thad: There is no general answer for that, it depends on the rolls. The general principle is that doubles being a great roll, you want to focus on offense. This is because the positions where both players anchor high are mostly about equal, while prime against prime positions are not.
But tactics often overrule that principle. 3-3 is one of the most difficult rolls. The 3 is not a great point, but making it unloads the stack on the 6. The 5 is great, but making it abandons the 8. The 10 is not great, but it unstacks the midpoint. The defensive 20 is great, but you would like to focus on attack.
After an opening 6-1, the value of advancing the back checkers goes up, so I like your move. At www.bgonline.org there are all rollouts for all second moves that are considered non-trivial, and 24/21(2) 13/10(2) indeed rates as the best move.

alanback: And most importantly, the FIBS formula is perfectly compatible with the BK formula. I told Fencer long ago that I was able to give him a formula that implemented the FIBS trick for taking into account the length of the match, while leaving the BK formula unchanged for single games. He seemed to be OK with the idea of changing the formula, but not willing to implement it any time soon.

It is not a backgammon-specific problem. The current rating system could also be exploited by playing multi-games chess matches against weaker players.

alanback: Terrific idea ! Turn off / turn on the cube would solve the autopass when owning the cube problem and do much more in the same time. There should be some red warning besides the board that says "you have turned off the cube in this game", though.

AlliumCepa: Not when you have the cube.

pgt: I agree with you, autopass (or the lack of it) is still the #1 weakness of BK.

Thad: I welcome the idea to play backgammon with different dices, but I wouldn't choose those ones. As playBunny pointed out, the fact that the sum of the two dices is the same as with ordinary dices is not very relevant in backgammon, and I don't like the fact that it would be impossible to full-prime the opponent (it would require 16 checkers).

I would welcome new dices such that
1) There are less doubles. This sounds nice.
2) There is a higher probability to hit a direct shot - which would favour the positional plays vs the racing plays.

A possibility would be one die ranging from 1 to 5, the other from 2 to 6. It would require some testing to tell how it would work.

Thad: "Unknown bug" is the best explanation I could think of.

alanback: Moreover, the slower you play, the better equity you have to collect points due to timeouts / resigns. It could lead people to play as slow as possible.
It could be a general rule in all tournaments that when somebody timeouts or resigns a backgammon, he or she automatically forfeits all games with maximal scores, even the already finished ones.

Andersp: If you win against someone who is 2200 it means... well, not much, since the BK rating formula is inaccurate for backgammon anyway !

AbigailII: Your logic is unfaulty ! However, things can be seen for another point of view : cubed backgammon IS backgammon, and single games are the degenerated case of matches to one point.
I take "degenerated" for its mathematical meaning, it is not supposed to be insulting to one-pointer lovers :-)

But of course you are basically right, there isn't a bigger difference between Triple and Single than between Cubed and Single.