Return Match for Komodo

[Jesse Gersenson]

All 4 games are here:
http://u.komodochess.com/games/neuman-rematch.pgn

[lkaufman]

All 4 games are here:
http://u.komodochess.com/games/neuman-rematch.pgn

I reviewed the fantastic final game, and it seems that none of the fireworks would have happened without Contempt. Komodo judged that it was better to be down a knight with decent but not adequate compensation than just to allow simplification two pawns down. One move before the end Black might still have held a draw with perfect play by taking on g6 with the pawn, sacrificing the Exchange, but of course perfect play is not possible for a human with limited time. It seems that Contempt is what makes these matches interesting.

[S.Taylor]

WhatI'm most looking forward to is when no GM can even DRAW to Komodo even when given knight odds, or a bishop odds. Or at most, the GM to get a rare draw. And a rare win at rook odds.

Well, not SUCH a rare win at rook odds, but certainly not automatic that a GM will win. ANY GM.

Odds of a rook and a piece, the GM should ALWAYS win , even 100% (100/0 of a hundred games), but only if the GM is fully focused the whole time.

Np sorry, i'll change that a little. A rook plus an exchange=7 point odds it should need to gurantee a top GM win every time. And that does not include if the GM is playing in simultaneous exhibition or a blitz game.

[Uri Blass]

WhatI'm most looking forward to is when no GM can even DRAW to Komodo even when given knight odds, or a bishop odds. Or at most, the GM to get a rare draw. And a rare win at rook odds.

Well, not SUCH a rare win at rook odds, but certainly not automatic that a GM will win. ANY GM.

Odds of a rook and a piece, the GM should ALWAYS win , even 100% (100/0 of a hundred games), but only if the GM is fully focused the whole time.

Np sorry, i'll change that a little. A rook plus an exchange=7 point odds it should need to gurantee a top GM win every time. And that does not include if the GM is playing in simultaneous exhibition or a blitz game.

I think that you are looking for the impossible.
I believe that a Knight is clearly too much handicap for a GM.

I also believe that humans can perform better at 2 pawns handicap and they simply suffer from lack of experience in these conditions.

I would like Fide to allow computers to play for fide rating with 2 pawn handicap and with knight handicap to see their real fide rating in these conditions.

[Frank Brenner]

In Game 4 Stockfish also plays the knight sac 11 Sxe6 with contempt = 0.

I dont believe that "contempt" was the feature that is responsible for this beautiful open tactical chess game (nr 4).

The reason was: luck

To have more exciting games like game 4 in the future in Computerchess Komodo has to learn the idea of raising the potential complexity of a position.

I want to explain what i mean with that:

Most dead draw positions are of very low (zero) potential of complexity.

The Starting position (no move played so far) has a high potential of complexity

A complicated position in the middle game that is game theoretic a draw, but white to move has to find the next 3 very sharp and only moves could be of moderate low potential complexity or of high potential complexity depending on the final position that appears on the board when white succesful finds the 3 right moves and plays the right path.

If white findes the 3 moves and the final position is a very much simplificated position the initial position is maybe of more or les low potential complexity. The potential complexity could be 3 or so on a interval between [0,20]

If the final position after finding the 3 moves is still complex, the initial position also have a very high potential complexity maybe 10 or so.


At the moment Komodo only deals with material. K trys to avoid exchanges ... but this feature is only a very low approximation of the goal to raise the potential complexity of a position.



The problem is: how to identify the potential complexity and how how to give it a number of 0....50 .. ?

Actually i have no real idea for that.

When i look at the live tcec games i see a lot of 0.00 zero evaluations.
It is hard find properties of those positions that identify this positions as positions of low potential complexity.
I think there is one other thing to deal with: making progress.

[lkaufman]

WhatI'm most looking forward to is when no GM can even DRAW to Komodo even when given knight odds, or a bishop odds. Or at most, the GM to get a rare draw. And a rare win at rook odds.

Well, not SUCH a rare win at rook odds, but certainly not automatic that a GM will win. ANY GM.

Odds of a rook and a piece, the GM should ALWAYS win , even 100% (100/0 of a hundred games), but only if the GM is fully focused the whole time.

Np sorry, i'll change that a little. A rook plus an exchange=7 point odds it should need to gurantee a top GM win every time. And that does not include if the GM is playing in simultaneous exhibition or a blitz game.

I think that you are looking for the impossible.
I believe that a Knight is clearly too much handicap for a GM.

I also believe that humans can perform better at 2 pawns handicap and they simply suffer from lack of experience in these conditions.

I would like Fide to allow computers to play for fide rating with 2 pawn handicap and with knight handicap to see their real fide rating in these conditions.

Well, it would be nice, but I think you can forget about FIDE in this regard. I think they are asking 20,000 dollars (or was it Euros?) for a computer to get a normal FIDE rating; I think they are overestimating its value by at least a factor of ten. But of course, if GMs got the standard 90' + 30" minimum time in these matches they would do better. I'm thinking that at some point we may try giving the GM this much time with maybe 1/3 of that for Komodo, along with material handicaps like f7 or Exchange or even two pawns. This would still keep the game length low enough to play two rounds a day and to ask for coverage by chess.com or others. That would remove the last excuse for the poor human results.
I agree with you Uri that knight odds will always favor a mid-level GM over any computer at 90' + 30". But I'm not certain of that. Knight odds is roughly double two pawn odds, though a lot depends on the choice of pawns. If we play another two pawn odds match, I'm thinking f2 plus b2,c2,d2, and e2 would be the next step up. We already know that c2 + f2 is really tough from the original Neuman match. Of course we can always go to two pawns plus first move if we need to increase it further. Playing first seems to be fairly important in these handicap games. It seems that in the opening, one pawn = two tempi (rather than the classical three) is about correct. So switching color is a half pawn swing.

[Laskos]

I agree with you Uri that knight odds will always favor a mid-level GM over any computer at 90' + 30". But I'm not certain of that.

?
Knight odds is definitely too much today for a mid-level GM at 90'+30'', and against a good GM it will be too much even with a perfect engine.

[Laskos]

I also believe that humans can perform better at 2 pawns handicap and they simply suffer from lack of experience in these conditions.

I would like Fide to allow computers to play for fide rating with 2 pawn handicap and with knight handicap to see their real fide rating in these conditions.

That would be interesting. I too believe that humans suffer from the lack of experience in these conditions. If the match was like the first one, 4 games from the same odds position, say e2f2, the outcome might have been a bit different.

[lkaufman]

I also believe that humans can perform better at 2 pawns handicap and they simply suffer from lack of experience in these conditions.

I would like Fide to allow computers to play for fide rating with 2 pawn handicap and with knight handicap to see their real fide rating in these conditions.

That would be interesting. I too believe that humans suffer from the lack of experience in these conditions. If the match was like the first one, 4 games from the same odds position, say e2f2, the outcome might have been a bit different.

yes, that's probably the main reason Neuman went from plus 3 at c2,f2 to minus 3 at e2 + b2,c2,d2, and f2. But it should be noted that two strong GMs, Movsessian and Lenderman, each got four tries at f7 handicap, but fared poorly, not winning a single game. Lenderman even had done some serious analysis of the handicap before the match. Some handicaps are much easier to prepare for than others.

[lkaufman]

In Game 4 Stockfish also plays the knight sac 11 Sxe6 with contempt = 0.

I dont believe that "contempt" was the feature that is responsible for this beautiful open tactical chess game (nr 4).

The reason was: luck

To have more exciting games like game 4 in the future in Computerchess Komodo has to learn the idea of raising the potential complexity of a position.

I want to explain what i mean with that:

Most dead draw positions are of very low (zero) potential of complexity.

The Starting position (no move played so far) has a high potential of complexity

A complicated position in the middle game that is game theoretic a draw, but white to move has to find the next 3 very sharp and only moves could be of moderate low potential complexity or of high potential complexity depending on the final position that appears on the board when white succesful finds the 3 right moves and plays the right path.

If white findes the 3 moves and the final position is a very much simplificated position the initial position is maybe of more or les low potential complexity. The potential complexity could be 3 or so on a interval between [0,20]

If the final position after finding the 3 moves is still complex, the initial position also have a very high potential complexity maybe 10 or so.


At the moment Komodo only deals with material. K trys to avoid exchanges ... but this feature is only a very low approximation of the goal to raise the potential complexity of a position.



The problem is: how to identify the potential complexity and how how to give it a number of 0....50 .. ?

Actually i have no real idea for that.

When i look at the live tcec games i see a lot of 0.00 zero evaluations.
It is hard find properties of those positions that identify this positions as positions of low potential complexity.
I think there is one other thing to deal with: making progress.

I agree with you, I've often thought about how to measure drawishness (opposite of complexity) in a position. It's not easy. The problem of progress is also critical, but there at least we took a step forward with Komodo 10.1. Regarding game 4, maybe the knight sacrifice wasn't so shocking, but the earlier Ng5 move setting it up was pretty hard to find. Did Stockfish play that? Komodo only did it with high Contempt.