Match at pawn odds: Rybka vs GM Jaan Ehlvest


     If one didn't know that "Rybka" is a computer program and not a human being, the obvious conclusion would be that Rybka fought with all her power until the match was won, but once there was no additional prize to win Rybka chose to go easy on her opponent, perhaps to lure him into a rematch or perhaps due to mental fatigue. However this would be attributing a degree of artificial intelligence to Rybka beyond anyone's imagination at this point. So we must seek alternate explanations.

     The final score of 5.5 - 2.5 was better than my forecast (4.5-3.5) and better than the median of the polls (between 4.5 and 5.0 for Rybka). On the other hand it was worse than the results that I would have predicted strictly based on computer-computer games. I assumed that the human player would adapt to the unusual circumstances whereas a computer would not, but perhaps I overestimated the degree of adaptation.

     But how is it possible that things could change so dramatically after the fifth game (I count the fourth as a success for Rybka, as Ehlvest felt he had lost all advantage after just a handful of moves and risked losing if he avoided repetition)? Ehlvest himself says that the problem was that he used too much time in the opening in the early rounds. Against human opponents it's not so terrible to have to average 10 seconds a move in the later stages, but against Rybka it was too difficult. This may be true, but I don't think it explains the dramatic difference between the two match parts.

     I think the explanation is that Ehlvest is a very classical player, likes to castle kingside, occupy the center, and play for positional advantages. When White is missing a kingside pawn, he can hope to use the opened file to attack the king. In game 1 (h2 off), the open file discouraged Black from castling short, and since the opening made castling long a bit cumbersome, Black left his king in the center and got crushed. In game 2 (g2 off) Black conceded space and postponed castling too long, perhaps keeping both options open, but a trick cost Black castling and the game. In game 3 both castled short and Black was a clean pawn up, but White still managed to get a winning attack on the king using the open "f" file. In game 5 (d2 off) Black returned the pawn right away to get attacking chances, but that merely led to a tactically complex game that favored the computer. In game 6(c2 off) Black played a classical defense (Slav) and after both sides castled short exploited his extra pawn in exemplary fashion. In game 7(b2 off) White chose a double fianchetto (generally not advisable) while Black played a classical Tarkakower setup, and soon Black had a big space advantage plus the pawn, and was apparently winning at a rather late stage of the endgame before going astray in time pressure and allowing a draw. The final (a2 off) game was basically quite nice for White, and Black only drew by heading to a pawn down but drawn rook endgame.


Conclusions: The human player should play classically, occupy the center and castle on the side where the computer is not missing a pawn (or kingside if it's a center pawn). It took Ehlvest a few games to figure this out. When the missing pawn is on the kingside, it's confusing for the human player, because there are very few normal openings in which Black castles queenside.

     The computer should avoid double fianchetto, and should make more use of pawns in the opening rather than just bringing pieces to good squares. For example, Ehlvest felt that with b2 off, White should play Bb2 and f4 to start, rather than just Bb2 and Nf3. Rybka needs to learn "shape", for example that knight or bishop directly behind a pawn is desirable. The opening play when out of book can be improved quite a bit, I feel.

     As for the handicaps, clearly the two edge pawn handicaps are pretty small, especially with the program playing White. This is due both to the immediate activity of the rook on the half-open file and to the natural inferiority of edge pawns, which can capture only one way. Probably they are just mildly bigger advantages than simply playing White in normal chess.

     The other six are not so different in magnitude. My computer-computer testing showed only a small variation in the size of these handicaps, except for "f2" being noticeably bigger than the others. The match suggests that for a human opponent, the kingside pawns are smaller handicaps, only because they require "strange" openings where Black should castle queenside. Ehlvest emphasized (after the match) the importance of the center in these games, and so felt that the knight's pawns should be smaller handicaps than the bishop's and center pawns as these pawns attack central squares. Based on all of this, it is entirely possible that it is not just a fluke that Ehlvest's win was with the "c" pawn handicap. Kingside pawns make for strange openings, center pawns give White immediate bishop and queen activity, and the "b" pawn can't help the center, so that leaves -- the "c" pawn!

      Actually I have always felt that the "c" pawn makes for the "cleanest" handicap -- play most like normal chess. If I'm right, then it's no surprise that Ehlvest won that game. To be fair, I should point out that the program "Fruit" did win a game with the c2 handicap from the late GM Alex Wojtkiewicz, though it was at a considerably faster time limit (25'+10").

     For the future, we can consider keeping the present format, but making the challenge more difficult for Rybka by increasing the increment, so that time management will be less important. Probably if the time limit had been 45'+30", the match would have been very close. Or if a shorter match is needed, perhaps just c,d,e,and f pawns would be chosen, although this would clearly be a larger handicap, perhaps more suited for a slightly lower rated GM. My personal goal is to see the day when Rybka can give the traditional "pawn and move" (Rybka plays black and removes f7) to a top GM in standard time control chess successfully, but we are a long way from that now as that handicap gives White an initiative, space, and better center after 1e4 due to tactical considerations. Even now this handicap is probably suitable for a marginal Grandmaster (2500 FIDE) at a rapid time control like 25'+10".


     Larry Kaufman