JJJ wrote: ↑Mon Mar 16, 2020 10:29 pm
I m happy Larry you re trying this match. Yes this time Komodo might finally be underdog. But maybe it won't last forever... I m like you, hoping one day the best engine could give a knight handicap against the strongest in the world.
Well, at tournament time control against top humans, this might be unachievable with current engines' paradigm, although the Komodo contempt makes first steps by a top engine to be more anti-human. But theoretically one can build a very anti-human engine, say one which creates 5-6 tactical puzzles 9-10 plies deep during each of the games. Even top humans might fall for 1-3 of these, the engine thus overcoming the heavy material handicap. Humans are very sensitive to material, give them only an overwhelming positional advantage, not material one, over Komodo, and even top humans might lose such a match.
So far the indications are that the use of Komodo MCTS rather than normal Komodo is a bigger factor against humans than Contempt, but this hasn't really been proven in a controlled experiment, it's just my opinion from playing it myself and from the few knight odds matches with masters we have held. You may be right about the material vs. positional advantage claim, but I'm not so sure. I was able to win a rapid game vs. Stockfish 11 (on 7 threads) the other day at pawn and three move odds (f7 off, e4 and d4 played, WTM), which is less than knight odds based on evals and on engine vs engine games. Pawn and three moves is mostly a positional handicap. It may depend on the engine.
I think that some future hybrid of NN, MCTS, and A/B may be really awesome at giving handicaps, but it doesn't exist yet.
In that handicap game pawn+3 moves, didn't you quickly get even more material advantage like 2 pawns and a very comfortable position? f7 pawn is big and affecting the King, an engine could give up a pawn to salvage the position. Also, SF has a simple contempt. I think it's more illuminating to play against Komodo from one of those 6-7 movers we researched 2-3 years ago.
I myself lose to Lc0 early versions faster than against latest SF. Early Lc0 had some intrinsic contempt, literally smacking me more in the vein of how I am beaten by a strong human. I am not sure what's the matter, MCTS search or NN or both.
Yes, I did indeed win a second pawn quite soon, with more to follow. I think that Stockfish played this game quite sub-optimally. So I guess by "positional advantage" you mean "positional advantage excluding king safety advantage". It does seem to be harder for humans to capitalize on advantages in development and piece placement than it does to take advantage of material or safer king. But for me king safety is in the positional category. It's just a matter of terminology.
What handicap do you need to be a fair match for either Komodo, Komodo MCTS or early (10,000 series) Lc0? I think you once estimated your level as 1800, if so I'm guessing a rook is almost but perhaps not quite enough? But maybe these engines already deteriorate too much down a rook, I'm not sure. I think our webmaster Jesse, who is around that Elo level, found rook handicap to be appropriate. But depends on time limit of course.
Well, it's SF being too serious when fearing for its king safety, if it "knew" that there is nothing serious about your threats, he wouldn't probably give up pawns to deflect them. That can be cured by a smarter contempt, say not to give up or exchange pawns, not to simplifying moves etc. Komodo does it somehow with say +150 contempt, right?
I was almost 1800 extrapolated to FIDE ratings some 15 years ago, when I was playing online. Today I am probably lower than 1600, and usually hate playing. Against regular Komodo I had 2-3 years ago reasonable chances of winning or drawing at rook + 1-2 takebacks handicap at 15m+10s with me sometimes overstepping the time limit by a little. Lc0 series 10,000 seemed nastier, especially on my first more serious GPU.
JJJ wrote: ↑Mon Mar 16, 2020 10:29 pm
I m happy Larry you re trying this match. Yes this time Komodo might finally be underdog. But maybe it won't last forever... I m like you, hoping one day the best engine could give a knight handicap against the strongest in the world.
Well, at tournament time control against top humans, this might be unachievable with current engines' paradigm, although the Komodo contempt makes first steps by a top engine to be more anti-human. But theoretically one can build a very anti-human engine, say one which creates 5-6 tactical puzzles 9-10 plies deep during each of the games. Even top humans might fall for 1-3 of these, the engine thus overcoming the heavy material handicap. Humans are very sensitive to material, give them only an overwhelming positional advantage, not material one, over Komodo, and even top humans might lose such a match.
So far the indications are that the use of Komodo MCTS rather than normal Komodo is a bigger factor against humans than Contempt, but this hasn't really been proven in a controlled experiment, it's just my opinion from playing it myself and from the few knight odds matches with masters we have held. You may be right about the material vs. positional advantage claim, but I'm not so sure. I was able to win a rapid game vs. Stockfish 11 (on 7 threads) the other day at pawn and three move odds (f7 off, e4 and d4 played, WTM), which is less than knight odds based on evals and on engine vs engine games. Pawn and three moves is mostly a positional handicap. It may depend on the engine.
I think that some future hybrid of NN, MCTS, and A/B may be really awesome at giving handicaps, but it doesn't exist yet.
In that handicap game pawn+3 moves, didn't you quickly get even more material advantage like 2 pawns and a very comfortable position? f7 pawn is big and affecting the King, an engine could give up a pawn to salvage the position. Also, SF has a simple contempt. I think it's more illuminating to play against Komodo from one of those 6-7 movers we researched 2-3 years ago.
I myself lose to Lc0 early versions faster than against latest SF. Early Lc0 had some intrinsic contempt, literally smacking me more in the vein of how I am beaten by a strong human. I am not sure what's the matter, MCTS search or NN or both.
Yes, I did indeed win a second pawn quite soon, with more to follow. I think that Stockfish played this game quite sub-optimally. So I guess by "positional advantage" you mean "positional advantage excluding king safety advantage". It does seem to be harder for humans to capitalize on advantages in development and piece placement than it does to take advantage of material or safer king. But for me king safety is in the positional category. It's just a matter of terminology.
What handicap do you need to be a fair match for either Komodo, Komodo MCTS or early (10,000 series) Lc0? I think you once estimated your level as 1800, if so I'm guessing a rook is almost but perhaps not quite enough? But maybe these engines already deteriorate too much down a rook, I'm not sure. I think our webmaster Jesse, who is around that Elo level, found rook handicap to be appropriate. But depends on time limit of course.
Well, it's SF being too serious when fearing for its king safety, if it "knew" that there is nothing serious about your threats, he wouldn't probably give up pawns to deflect them. That can be cured by a smarter contempt, say not to give up or exchange pawns, not to simplifying moves etc. Komodo does it somehow with say +150 contempt, right?
I was almost 1800 extrapolated to FIDE ratings some 15 years ago, when I was playing online. Today I am probably lower than 1600, and usually hate playing. Against regular Komodo I had 2-3 years ago reasonable chances of winning or drawing at rook + 1-2 takebacks handicap at 15m+10s with me sometimes overstepping the time limit by a little. Lc0 series 10,000 seemed nastier, especially on my first more serious GPU.
Komodo does have a "smarter" Contempt than SF (meaning it has more consequences), and SF limits the value to 100 while we allow up to 250, but there is nothing in either Komodo or Stockfish that would discourage giving up a pawn to avoid bad things happening to the king. I suspect that the automatic Contempt in SF (even with Contempt set to 0) has some bad side effects that might explain this behavior. Yes, Ray and I and also GM Naroditsky all found that the Lc0 10,000 series was really good at handicap play; too bad all the later ones were awful at it.
lkaufman wrote: ↑Tue Mar 17, 2020 6:15 pm
there is nothing in either Komodo or Stockfish that would discourage giving up a pawn to avoid bad things happening to the king
I wonder how would one even add such a thing to the engine. If the engine sees that there's a line of moves that leaves the king in a compromised position leading to a lost game, how is contempt supposed to ignore that? And even if it could, if Contempt is supposed to avoid draws to give more wins to the engine, how can we know that the human will not find that attacking line and play it so that'd just decrease the engine's performance?
A pawn seems like a cheap price to avoid that entire problem.
lkaufman wrote: ↑Tue Mar 17, 2020 6:15 pm
there is nothing in either Komodo or Stockfish that would discourage giving up a pawn to avoid bad things happening to the king
I wonder how would one even add such a thing to the engine. If the engine sees that there's a line of moves that leaves the king in a compromised position leading to a lost game, how is contempt supposed to ignore that? And even if it could, if Contempt is supposed to avoid draws to give more wins to the engine, how can we know that the human will not find that attacking line and play it so that'd just decrease the engine's performance?
A pawn seems like a cheap price to avoid that entire problem.
Well, we could simply make Komodo more materialistic with higher Contempt, for example by lowering Dynamism. But as you say, it's not clear that this would actually be beneficial, although if I had to guess I would think it might help, because humans are better at converting a material advantage than at working out a complicated, narrow path to a mating attack. It's not meaningful to talk about a pawn being a cheap price to avoid a problem, it just depends on how big the problem is. In the game in question, it seemed like a bad practical decision to me, even if it was correct against another Stockfish level opponent.
MikeGL wrote: ↑Mon Mar 16, 2020 10:49 am
Sounds exciting. Thanks for the heads up.
Ballpark score could be 3.5 - 2.5 win for Komodo.
All draws plus 1 win for the engine.
Well, as Tal reportedly said when told that Bobby Fischer said in the 1960s that he could give knight odds to any woman chessplayer, "Fischer is Fischer, but a knight is a knight" (or something like that). I believe that eventually Komodo will win a match like this, but not this year, we need a major breakthru for that to happen in my opinion. We did beat MVL in a (fast) rapid match at handicaps averaging about a 2.0 eval for him, but knight odds is way more than that. In the opening, a knight is worth nearly four pawns, so even with the first move, White is down about 3.5.
Very interesting.
2 GM wins, 4 draws. In this range, I guess. In other metric, something like 80-90% range that this young active GM won't lose the match.
Having played some test games of Komodo vs. myself and my son Ray under similar conditions, I'm now inclined to agree with your forecast of 2 GM wins and 4 draws. I think it performs about as well as a FIDE 2400 would with no handicap.
This will be exciting to watch!
I'm going to stick with my prediction that Komodo will win comfortably.
I apologize if these questions have already been answered, but
1. Which version of Komodo will be playing? Will GM Smerdon be told which version he's playing against?
2. What will Komodo's contempt settings be, and will Gm Smerdon know them?
Thank you, GM Kaufman.
