LearningLemming CCT10 report

[BubbaTough]

Howdy,

I have been requested by a couple folks to give a LearningLemming CCT10 (http://www.cctchess.info/) report, so here it is.

My hope for the tournament was twofold: consistent play against non-elite engines, and to get out of every opening with an OK position. The second goal was in response to DIEP's mixed complement during our recent match (http://taccl.org/ACCATeamMatch/2008ACCA ... Match.html) where he said I had a world class engine EXCEPT for its opening.

LL notes: ran on 1 processor (2.66 core 2 duo) using a book developed solely using automated learning techniques.

(0-0)Game 1. Galurung vs. LL. LearningLemming was seeded 25th based on its poor performance last year, and had to face an 8 processor Glaurung as black first round. The opening was a bit off beat and very complicated, but LL achieved a nice advantage and by the 16th move the Glaurung operator was beginning to fear a loss. Glaurung defended well, and transposed into a slightly inferior endgame which was drawn on move 116. The result pleased both of us, me because of a draw against a very good engine, and the Glaurung operator for surviving after a tough opening.

(0.5-0.5)Game 2. LL vs. Averno. We left book at move 13 with a complicated but balanced position. Averno ruined LearningLemming's kingside structure and grabbed a pawn, in exchange LL got the 2 bishops and an attack. Averno eventually had to escape a vicious assault on its king by transposing into a lost endgame, which LearningLemming managed to win despite a few scares.

(1.5-0.5)Game 3. Petir vs. LL. I used to test some against Petir, so knew this would be a tough match. LL book went fairly deep into a ruy lopez with a solid position, but proceeded to make a number of slight positional errors that had me quite concerned. With the queenside locked both engines agreed to rip open the kingside pawns in front of both kings, and Lemming got very active in exchange for some pawn weaknesses, and outplayed Petir to achieve an easily won good knight vs. bad bishop endgame.

(2.5-0.5)Game 4. LL vs. Rybka. I would rather play black vs. Rybka, because Rybka books seem to allow LearningLemming an easy equal position (in exchange for keeping a lot of pieces on). Thus, playing white seems a waste. Anyway, both engines left book early with an equal position, and went into a mildly complicated (but somewhat drawish) endgame. Amusingly enough, while Vas and I were busy chatting about techniques for improving engine understanding of contempt, LearningLemming was showing its contempt for Rybka, pushing very hard for a win in a drawish position, and achieving a losing position. We went into an endgame where LL had 4P+B+N vs. 6P+R and Rybka easily outplayed LL for the win.

(2.5-1.5) Game 5. Crash Test Dummy vs. LL. This was probably my favorite game. We left book around move 8 with a dynamic endgame, and LL played some very interesting and risky moves, offering to sacrifice material in multiple different ways to try to trap or discombobulate opponent pieces. CTD decided not to win the exchange and trap its knight on the kingside, but won a pawn instead in such a way that its other knight got stuck on the queenside with its escape path guarded by pieces instead of pawns, a situation most programs do not have patterns to detect (unlike traditional knight stuck on A8/H8 situations) but humans would be wary of. Eventually LL picked up the knight and the victory, in a game that reminded me of a classic human beats computer style game.

(3.5-1.5) Game 6. LearningLemming vs. E.T. Chess. This was an interesting match, in that almost all other opponents were running on 4 or 8 processors, while E.T. Chess and LearningLemming were competing with the big boys using only 1. The opening went quickly to a drawish endgame, but then LL allowed a strong looking (to most engines) outside passed pawn in a double rook endgame in exchange for disconnecting and discombobulating the opponent rooks. The position looked won for LL to me, but after tying down E.T.'s pieces LL started shuffling aimlessly instead of forcing a passed pawn down E. T.'s throat, and let the position slip into a drawn position. It then took some crazy risks trying to win, and eventually drew. A disappointing result, as a win would have put me in a tie for 2nd with Rybka, Zappa, Naum, and Glaurung going into the last round.

(4 - 2) Game 7. Junior vs. LL. Junior was running on 16 processors, and was the only engine to search significantly deeper than LL. The opening went well, with LearningLemming achieving a solid position and a big edge on the clock by the time it was out of "book". Junior generated a nice attack, and crushed LL in a VERY complicated game where both kings were exposed but junior was never in any real danger.

Overall, I was very happy with the tournament result. My only losses were against (world / x-world champions {Rybka and Junior}) and Lemming's self-generated book held its own against the professionals. This was a solid reinforcement of the results at the ACCA closed championships (http://taccl.org/ACCA2007Championships/ ... sults.html) where LL finished 3rd, losing only the Rybka and Naum. Lemming's aggressive pruning allowed it to compete with strong multithreaded engines without some of the inconsistencies some of the previous versions of LearningLemming have displayed. I guess the next step for LearningLemming should be to add multi-processor support, so that it can compete with the big-boys on a more equal footing.

-Sam

[Alessandro Scotti]

Thanks for the report, I enjoyed it a lot!

I was not aware of Learning Lemming, it's very nice to see a strong engine with new ideas. Besides, lemmings are almost as cute as my hamsters!

[Werner]

Thanks Sam for the nice report.
I watched CCT10 games here and I think you made a very interesting engine.
I wish you good luck for implementing mp features.

regards

[Marc Lacrosse]

Nice report, Sam !

Do you intend to release a public version of LearningLemming ?

Marc

[BubbaTough]

Nice report, Sam !

Do you intend to release a public version of LearningLemming ?

Marc

I plan to release a public version, but I have no idea when. I doubt it will be anytime soon. I would prefer not to release a version until it is reasonably bug-free, which may take a while since my work on it is pretty intermittent.

-Sam