can chess program see win for white after e4 h6 d4 a5?

[lkaufman]

Okay, I mistakenly thought you meant a series of games by one player, in which case he would know the opening after the first game. Still, I think this only makes a small difference, maybe I would raise my 60% estimate to 65% or so. Milov knew the handicap in advance, but did not seem to make good use of this knowledge as he could not anticipate what I put into the book. He won the match not in the openings, but just by having the technique to win a won game (at least in the two games out of the six handicap games that he did win, the others being drawn). In your hypothesized match, the GM technique should be good enough to draw the majority of the great positions he would get out of the opening, though his winning chances would be small. I mean if Milov did not lose any of the six handicap games, why should another 2700 player lose half or more of the games at the modestly reduced handicap you propose?

[Milos]

I mean if Milov did not lose any of the six handicap games, why should another 2700 player lose half or more of the games at the modestly reduced handicap you propose?

Because the advantage/handicap is really much smaller than what it looks at the first look.
Dann tried with Stockfish, I tried with Rybka and Firebird playing all against itself the very same oppening, and when duplicate games were removed it was always around 55% for white, not more. Now if you think that strong GM against engine can get 50% on average that would mean the strongest engines are only 35 elo stronger than strongest GM's. That simply doesn't hold.

[lkaufman]

These results are bizarre! You are saying that in engine play White's advantage against two stupid Black moves is not appreciably more than in the opening position!!? This is completely contrary to my experience with computer chess, which shows that computers are quite good at winning or at least scoring heavily from obviously favorable positions. If your results are really statistically valid it is a sad commentary on computer chess; are programs really that weak that they cannot exploit such a huge advantage? I'll try some game like this myself with some engines.

[Milos]

These results are bizarre! You are saying that in engine play White's advantage against two stupid Black moves is not appreciably more than in the opening position!!? This is completely contrary to my experience with computer chess, which shows that computers are quite good at winning or at least scoring heavily from obviously favorable positions. If your results are really statistically valid it is a sad commentary on computer chess; are programs really that weak that they cannot exploit such a huge advantage? I'll try some game like this myself with some engines.

I say just that this kind of opening was never deeply investigated since it's assumed too weak, it does not necessarily mean it's that weak.
Moreover, in CC advantage of white against black is far less than 10 percentage points. In the balanced set of opening positions it should not be more than 5 points.

[lkaufman]

What size samples are you talking about when you say White only scored 55% in this 1e4 h6 2d5 a5 position? I've been running Monte-Carlo games in Rybka 3 at seven ply (which we all know really means ten ply), and after a thousand games White has scored 71.2%, or 157 Elo. This is about what I would expect for a position that is a borderline win. White's advantage is over four times the 55-45 margin in normal chess (GM games), which is quite consistent with my estimate of White's advantage as two tempi, since the initial position is half a tempo better for White.
As to the results in balanced openings, this only shows that balanced openings are chosen to be more equal than the results of best opening play. It is well known that in top GM play, Black has to suffer in the opening and is normally happy with a draw. I don't know much about CC, but if the results are as close as you say it may either indicate that games of weak players are included, or that players with White chose exciting but inferior openings like gambits because they are not playing for money, only for enjoyment.

[Milos]

What size samples are you talking about when you say White only scored 55% in this 1e4 h6 2d5 a5 position? I've been running Monte-Carlo games in Rybka 3 at seven ply (which we all know really means ten ply), and after a thousand games White has scored 71.2%, or 157 Elo. This is about what I would expect for a position that is a borderline win. White's advantage is over four times the 55-45 margin in normal chess (GM games), which is quite consistent with my estimate of White's advantage as two tempi, since the initial position is half a tempo better for White.
As to the results in balanced openings, this only shows that balanced openings are chosen to be more equal than the results of best opening play. It is well known that in top GM play, Black has to suffer in the opening and is normally happy with a draw. I don't know much about CC, but if the results are as close as you say it may either indicate that games of weak players are included, or that players with White chose exciting but inferior openings like gambits because they are not playing for money, only for enjoyment.

10 ply is too shallow (too much blunders), also remove all the duplicates. Moreover, fixed ply games are bad for testing for numerous reasons (too low variety as the most important). With Rybka I had 100 games sample in 3'+3'', on 2 cores 2.1 GHz machine and after removing duplicates it was +22/-10/=68.
Regarding white/black win ratio it's always with either test sets like Noomen, Sedat, etc., or standard books truncated to 12 moves, alternating colors and top rated engines.

[lkaufman]

I also ran a thousand games the same way with the c2 pawn missing from the start position, and Black scored 68.5% or +135 Elo. Thus based on these two tests, the handicap proposed in the topic is actually a bit larger than the c2 pawn!

[lkaufman]

"10 ply is too shallow (too much blunders), also remove all the duplicates. Moreover, fixed ply games are bad for testing for numerous reasons (too low variety as the most important). With Rybka I had 100 games sample in 3'+3'', on 2 cores 2.1 GHz machine and after removing duplicates it was +22/-10/=68.
Regarding white/black win ratio it's always with either test sets like Noomen, Sedat, etc., or standard books truncated to 12 moves, alternating colors and top rated engines.

"

I estimate that ten ply searches play somewhere around the level of human IMs in tournaments, so "too many blunders" seems a bit harsh by human standards. In any case I'd much rather trust a thousand ten ply games than a hundred blitz games. Perhaps you are unfamiliar with the MC feature of Rybka 3; there should be no duplicates, and the variety should be quite good. If you are simply running games relying on the nondeterministic nature of MP to get some variety, your sample will be vastly less varied than mine. Most likely, you are getting many games that are identical for several moves, and probably White has already made one or more poor moves in those moves. For example, Rybka loves to play Nc3 and Nf3 very early, in front of the c2 and f2 pawns, but in this case this is probably poor strategy. The MC method insures that a wide variety of opening schemes will be tried.
Regarding the standard books, I think that they exclude openings which give White too large an advantage, so while they are fine for evaluating which engine is best, they understate White's true advantage.

[Dann Corbit]

This opening (because of the edge pawn moves) reminds me of Yasser Seirawan's "Winning Chess Openings" story called "Yasser Seirawan verses an Unknown Friendly" starting on page 5.

[muxecoid]

What is the rating increase given by good opening book? Suppose that we have the same engine playing as black either with no opening book at all or with h6, a5 as the only opening. Also suppose that the white player has a rather strong book.

Against an engine with a good general purpose book is it better to play h6, a5 instead of having no opening book at all?