Report from the 8th Computer Olympiad, Graz

Achim have submitted a report for the 8th Computer Olympiad held in Graz. The article shows some pictures and interesting some positions.

Programmers and maintainers of Chess, Poker, Backgammon, Checkers and a few other games went to Graz, Austria to take part in the 8th Computer Olympiad and Computer Chess World Championship from 23rd to 27th of November.DomRound about 60 programs in all categories played in impressing surroundings. Computer Chess was placed in a futuristic cave deep inside a hill, Backgammon and Go was played in the Casino Graz. The organization of this event was outstanding, Austrian television, the major of Graz and local politicians were present at the opening ceremony. We got an bulletin every day, there was a fast internet connection, and, and, and ...

Unfortunately only two Backgammon bots had the heart to come: BGBlitz, an underestimated but very strong java engine written by Frank Berger and GNU Backgammon, "coached" by Achim Mueller. As last year in Maastricht, Netherlands, other bots did not participate. Neither Snowie, nor Jellyfish, mloner, motif or other "elder" statesmen of backgammon did show up.

BGBlitz competed with an enhanced version of 1.6. Inside was a new engine and it was now possible, to let the bot evaluate 3ply. Because 3ply of BGBlitz was much slower than gnubg's 2ply (which is actually also 3ply in the "non-unix" world), Frank Berger built in an algorithm which let the engine switch back to 2ply automatically in non-contact positions. GNU Backgammon used the engine 0.14, developed by Joseph Heled in December 2002.

Both players agreed on the following modus:Achim and Frank Best of five 13-point matches, settings up to 3ply (respective 2ply for GNU Backgammon), self made bearoff databases allowed. At least three matches should be played on Sunday, if necessary, a fourth and fifth match on Monday. For generating rolls we used Franks BGBlitz, seed was determined by the serial number of my mouse. And this was my first error in the event. Never agree on your opponents dice unsolicited!

The contest turned out to be a real battle over five matches with a thrilling final. A quick analysis by Snowie proved, that both computers played on a very high level (if Snowie is the leveling rule at all). BGBlitz scored a 2.064 world class and GNU Backgammon a 1.112 extra-terrestrial over all five matches.

GNU Backgammon won the first match very easily. Final score was 13-4. The huge difference was - of course - determined by a certain luck factor, but also by timid cube decisions made by BGBlitz. The Java-Bot dropped at least three takeable doubles. The most incredible drop is shown here, other will follow in the deeper analysis.

1st match: GNU Backgammon 9, BGBlitz 4 (match to 13 points)
GNU Backgammon doubles to 2
|1X '1X2X1X3X2O4X ' ' '3O|
|   |
| [2]  |      |
| '2O2O1O1O2O|   |1O1O ' ' '3X|

Pipcounts: GNU Backgammon 106, BGBlitz 121
Position ID: tlrBwQBpnQcHAA Match ID: MBmgAZAAIAAA

- BGBlitz rejects

In this match I made my second decisive error (yes, even when bots play it's possible to involve human insufficiency). I gave Frank information about GNU's opinion on the questionable cube decisions. Something you would never do in "real life play" against human opponents in tournaments.

After this depressing experience Frank changed the doubling strategy of BGBlitz from "normal" to "aggressive." Indeed you could see an enhancement from then on. There were still some discussable cube decisions, but you could also see, that the "blitz" was biting harder. The second match also went to GNU Backgammon (13-10), but it was much closer. In this match we heavily discussed a running game pass by BGBlitz at the score of 11-9.

2nd match: GNU Backgammon 11, BGBlitz 8 (match to 13 points)
GNU Backgammon doubles to 2

|2X2X '2X3X3X1X ' ' ' '2X|
|   |
| [2]  |      |
| ' '2O3O2O3O|   |2O2O ' ' '1O|

Pipcounts: GNU Backgammon 78, BGBlitz 88
Position ID: 7HYbAgCb3QUDAA Match ID: MBmgAbAAQAAA

- BGBlitz rejects

And here came error number two. What would you normally do being an 87,5% favorite in the competition? Graz Right, when you have a run, then run. What did stupid Achim do? I suggested a break for a very late lunch and some fresh air. We went out of the casino for some tasty "Strudel" and "Glühwein" at the market place next to the town hall. Half an hour later we returned - very relaxed - to the battlefield to play the third (and what I hoped last) match.

Far from it! Somehow BGBlitz did also recreate with some fresh electric power or flavorsome bits and bytes. Or the German coffee machine did somehow communicate with his creators mobile phone reading a SMS sent by Frank's family: "Death to all GNU's. We all cross fingers here at home." The match itself was not that interesting. BGBlitz won 13-6 and Frank did a sigh of relief not getting the maximum penalty.

We still had some spare time, so we decided to start with match 4. The first game begun with a strange opening error by GNU Backgammon:

4th match: GNU Backgammon 0, BGBlitz 0 (match to 13 points)
GNU Backgammon to play 33

|2O ' ' ' '5X '3X ' ' '3O|
|   |
|  33  |      |
|2X ' ' ' '5O|   | '3O1O1O '5X|

Pip counts: GNU Backgammon 167, BGBlitz 160

- GNU Backgammon moves 24/21(2) 6/3(2)
1224/21(2) 6/3(2) 51.16%
 0.567 0.152 0.007 - 0.433 0.093 0.002 
 2224/15* 13/10 51.16% ( -0.00%)
 0.570 0.171 0.012 - 0.430 0.124 0.004 
 3224/15* 24/21 51.15% ( -0.01%)
 0.571 0.160 0.012 - 0.429 0.120 0.004 
 4224/21(2) 8/5(2) 51.14% ( -0.02%)
 0.562 0.162 0.008 - 0.438 0.097 0.003 
 5224/21(2) 13/10(2) 51.12% ( -0.04%)
 0.565 0.149 0.007 - 0.435 0.094 0.002 

The match went on and at the score of 5-3 BGBlitz offered an very interesting double.

4th match: GNU Backgammon 3, BGBlitz 5 (match to 13 points)
BGBlitz on roll, cube decision?

|1X '2O1O2X4X2X ' '2O2X '|
|1 |
|      | [2]  |
|1X2O '2O2O2O|   | ' '2O ' '2X|

Pip counts: GNU Backgammon 146, BGBlitz 147

- BGBlitz doubles

GNU judged this as a "close double/take" and finally lost a gammon. After the game I asked for a time out. Schlossberg GNU Backgammon seemed to be very groggy, and we agreed on adjourning to Monday morning.

The next morning we moved to the impressing playing ground in the big cave. Parallel to our matches the "Advances in Computer Games Conference" took place including some very interesting presentations on endgame bases in chess, search algorithms in games and so on.

The final of the fourth match was a cranky 1-2-3-backgame by GNU Backgammon. Unfortunately I missed to save this game, but I remember one position shown at the end of the paragraph. Anyhow it was a shame, because this game would have shown some real weaknesses of both bots. The ending was stamped by some unspectacular doubles of BGBlitz and two missed shots by GNU Backgammon, loosing a gammon and the match 3-13.

4th match: GNU Backgammon 3, BGBlitz 9 (match to 13 points)
GNU Backgammon to play 22
| '1X1X '1X2X ' ' ' ' '1O|
|   |
|  22  |      |
|2X2X4X3O3O4O|   |2O2O ' '2X '|

Pipcounts: GNU Backgammon 232, BGBlitz 94
Position ID: uHsbBABKAwzwNg Match ID: AQGpATAASAAA

Guess, what GNU Backgammon did play here ... Yes, it was 14/12*/10, 14/12 and ... 6/4!

DomSo it was all down to one last match. I was more worried about Frank's health then about winning the match. He did never participate in a "human" tournament and looked like he could hardly stand the stress. GNU Backgammon had a good start into the match and was leading 7-2 and 9-3. But with a risky double followed by a gammon win BGBlitz was back and leading 10-9. Trailing 11-9 GNU Backgammon offered a perfect and instructive match score double in an opening position with some gammon chances.

5th match: GNU Backgammon 9, BGBlitz 11 (match to 13 points)
GNU Backgammon on roll, cube decision?

|1O1O ' '3X4X '2X ' ' '2O|
|   |
| [2]  |      |
|2X ' '2O '4O|   | '3O ' '2O4X|

Pip counts: GNU Backgammon 155, BGBlitz 151
Position ID: mOeYAShwz+ABMA Match ID: MBmgAZAAWAAA

- GNU Backgammon doubles

A few rolls later BGBlitz found itself with 2 checkers on the bar against a 5-point board, but managed to come back and finally win the game, match and title. Congratualion to Frank, who Siegerehrung works now for more than 6 years on his BGBlitz and finally was rewarded with the second title after Maastricht in 2002.

Though GNU Backgammon lost, the trip to Graz was a very interesting experience again. Graz is a very nice city, we had a lot of fun there and really regretted that we could stay only for 3 days. Coaching a bot in a "face to face" competition is totally different from doing this at home and almost so thrilling as a "human" match in a big tournament.

posted at 18:26:32 on 09/26/04 by webadmin - Category: Reports