How powerful is the Queen compared to 2 Rooks ?

[lkaufman]

For one, there is nothing wrong with a score of +1.53 and failing to win: the draw margin in Chess is about 1.5 Pawn, and +1.53 thus means the position offers about a 50-50 chance on win / draw.

I was very surprised to read this sentence. All of my tests in past years put the 50/50 win/draw line at about 3/4 of a pawn, not 1.5 pawns. Of course it depends a bit on which engine you are talking about and time limits, etc., but for Komodo (and Houdini too I think) 3/4 pawn is about right. A "clean" pawn up is evaluated as somewhat less than 1.0 by Komodo in the opening and somewhat more in the endgame, but on average 1.00 should be pretty close to accurate. The recent SF versions have very inflated scores (to say nothing of Lc0) which report a neutral pawn plus as more than 1.0, so if yu are basing this on recent SF that could be part of the problem, but even with SF I don't think the 50/50 line would be much over 1.0. So I'm asking what your source for the 1.5 pawn claim was; perhaps it was based on some very weak engine, or one that reports unrealistic scores for a pawn plus?

It was just a rule of thumb I used when I was still playing chess myself: being a clean Pawn ahead is usually not enough to win the game. (Of course I was just a patzer compared to you!) Most KRPKR are draw, with minors instead of Rooks it is even a dead draw. And in these cases the Pawn even is a passer, which should put it on the high end of the Pawn-value spectrum. Even KPK is often a draw, although without pieces advatages count much heavier.

My experience with weak engines (Fairy-Max ~2000 Elo) is that classical Pawn odds results in a 68% score (color-averaged). This doesn't seem to depend much on the level of play (as changed by varying the TC). I once saw a posting on the Rybka forum where someone had measured this with Rybka to be 70-72% (but I was not sure whether this was color averaged). Note that a 50-50 win:draw would be a 75% score, so a draw margin > 1 Pawn seems even to hold in the opening.

I don't know anythinhg about Komodo's evaluation scale; I don't have Komodo. More importantly, the 1.53 score was ascribed to LC0, so it really depends on how LC0 translates winning prospects into centi-Pawn. Which is just a matter of presentation, and has nothing to do with the quality of play: any monotonous transformation of the score would lead to exactly the same play.

Evaluating the given position as 0.00 seems a mistake: it means the engine will give up trying, even against a much weaker opponent. No matter how large a contempt you set. The contempt would only help to avoid immediate draws, but it would not help to distinguish between lines that keep the opponent under pressure, and those that blunder away one or two Pawns.

What would be Komodo's static evaluation of that position, btw?

High level (human or engine) chess is mostly about trying to win a pawn and then promote it. Grandmasters clearly believe that in general an extra pawn should win if there is no compensation or drawing factors; if this were not so chess would be pretty much unplayably drawish. But the win/draw line does seem to be uncomfortably close to a pawn, you don't need much comp to be able to hold with perfect play. Recently chess.com ran a series of games between top engines (alternating sides) with the f2 pawn removed from the start position, and Black won about half the games, with the others drawn except for one White win (out of about 30 games). With f2 gone the only compensation for the pawn is the first move, but this is fairly significant; f7 handicap is clearly lost for Black based both on analysis and on actual results of top engines.
Regarding the specific position in the thread, Komodo gives it a zero or near-zero score at low depths, which rises starting around depth 10 to a max of about half a pawn, which drops back to zero when reaching depth 25. Komodo MCTS gives a score around a quarter pawn after a second or so, which gradually declines asymptotically approaching zero after a while.

[Chessqueen]

For one, there is nothing wrong with a score of +1.53 and failing to win: the draw margin in Chess is about 1.5 Pawn, and +1.53 thus means the position offers about a 50-50 chance on win / draw.

I was very surprised to read this sentence. All of my tests in past years put the 50/50 win/draw line at about 3/4 of a pawn, not 1.5 pawns. Of course it depends a bit on which engine you are talking about and time limits, etc., but for Komodo (and Houdini too I think) 3/4 pawn is about right. A "clean" pawn up is evaluated as somewhat less than 1.0 by Komodo in the opening and somewhat more in the endgame, but on average 1.00 should be pretty close to accurate. The recent SF versions have very inflated scores (to say nothing of Lc0) which report a neutral pawn plus as more than 1.0, so if yu are basing this on recent SF that could be part of the problem, but even with SF I don't think the 50/50 line would be much over 1.0. So I'm asking what your source for the 1.5 pawn claim was; perhaps it was based on some very weak engine, or one that reports unrealistic scores for a pawn plus?

It was just a rule of thumb I used when I was still playing chess myself: being a clean Pawn ahead is usually not enough to win the game. (Of course I was just a patzer compared to you!) Most KRPKR are draw, with minors instead of Rooks it is even a dead draw. And in these cases the Pawn even is a passer, which should put it on the high end of the Pawn-value spectrum. Even KPK is often a draw, although without pieces advatages count much heavier.

My experience with weak engines (Fairy-Max ~2000 Elo) is that classical Pawn odds results in a 68% score (color-averaged). This doesn't seem to depend much on the level of play (as changed by varying the TC). I once saw a posting on the Rybka forum where someone had measured this with Rybka to be 70-72% (but I was not sure whether this was color averaged). Note that a 50-50 win:draw would be a 75% score, so a draw margin > 1 Pawn seems even to hold in the opening.

I don't know anythinhg about Komodo's evaluation scale; I don't have Komodo. More importantly, the 1.53 score was ascribed to LC0, so it really depends on how LC0 translates winning prospects into centi-Pawn. Which is just a matter of presentation, and has nothing to do with the quality of play: any monotonous transformation of the score would lead to exactly the same play.

Evaluating the given position as 0.00 seems a mistake: it means the engine will give up trying, even against a much weaker opponent. No matter how large a contempt you set. The contempt would only help to avoid immediate draws, but it would not help to distinguish between lines that keep the opponent under pressure, and those that blunder away one or two Pawns.

What would be Komodo's static evaluation of that position, btw?

High level (human or engine) chess is mostly about trying to win a pawn and then promote it. Grandmasters clearly believe that in general an extra pawn should win if there is no compensation or drawing factors; if this were not so chess would be pretty much unplayably drawish. But the win/draw line does seem to be uncomfortably close to a pawn, you don't need much comp to be able to hold with perfect play. Recently chess.com ran a series of games between top engines (alternating sides) with the f2 pawn removed from the start position, and Black won about half the games, with the others drawn except for one White win (out of about 30 games). With f2 gone the only compensation for the pawn is the first move, but this is fairly significant; f7 handicap is clearly lost for Black based both on analysis and on actual results of top engines.
Regarding the specific position in the thread, Komodo gives it a zero or near-zero score at low depths, which rises starting around depth 10 to a max of about half a pawn, which drops back to zero when reaching depth 25. Komodo MCTS gives a score around a quarter pawn after a second or so, which gradually declines asymptotically approaching zero after a while.

Do you believe that Komodo can become a stronger engine than Alpha Zero if you develop a combination of Neural Networks combined with either Alpha-beta Search or Pruning?

[lkaufman]

For one, there is nothing wrong with a score of +1.53 and failing to win: the draw margin in Chess is about 1.5 Pawn, and +1.53 thus means the position offers about a 50-50 chance on win / draw.

I was very surprised to read this sentence. All of my tests in past years put the 50/50 win/draw line at about 3/4 of a pawn, not 1.5 pawns. Of course it depends a bit on which engine you are talking about and time limits, etc., but for Komodo (and Houdini too I think) 3/4 pawn is about right. A "clean" pawn up is evaluated as somewhat less than 1.0 by Komodo in the opening and somewhat more in the endgame, but on average 1.00 should be pretty close to accurate. The recent SF versions have very inflated scores (to say nothing of Lc0) which report a neutral pawn plus as more than 1.0, so if yu are basing this on recent SF that could be part of the problem, but even with SF I don't think the 50/50 line would be much over 1.0. So I'm asking what your source for the 1.5 pawn claim was; perhaps it was based on some very weak engine, or one that reports unrealistic scores for a pawn plus?

It was just a rule of thumb I used when I was still playing chess myself: being a clean Pawn ahead is usually not enough to win the game. (Of course I was just a patzer compared to you!) Most KRPKR are draw, with minors instead of Rooks it is even a dead draw. And in these cases the Pawn even is a passer, which should put it on the high end of the Pawn-value spectrum. Even KPK is often a draw, although without pieces advatages count much heavier.

My experience with weak engines (Fairy-Max ~2000 Elo) is that classical Pawn odds results in a 68% score (color-averaged). This doesn't seem to depend much on the level of play (as changed by varying the TC). I once saw a posting on the Rybka forum where someone had measured this with Rybka to be 70-72% (but I was not sure whether this was color averaged). Note that a 50-50 win:draw would be a 75% score, so a draw margin > 1 Pawn seems even to hold in the opening.

I don't know anythinhg about Komodo's evaluation scale; I don't have Komodo. More importantly, the 1.53 score was ascribed to LC0, so it really depends on how LC0 translates winning prospects into centi-Pawn. Which is just a matter of presentation, and has nothing to do with the quality of play: any monotonous transformation of the score would lead to exactly the same play.

Evaluating the given position as 0.00 seems a mistake: it means the engine will give up trying, even against a much weaker opponent. No matter how large a contempt you set. The contempt would only help to avoid immediate draws, but it would not help to distinguish between lines that keep the opponent under pressure, and those that blunder away one or two Pawns.

What would be Komodo's static evaluation of that position, btw?

High level (human or engine) chess is mostly about trying to win a pawn and then promote it. Grandmasters clearly believe that in general an extra pawn should win if there is no compensation or drawing factors; if this were not so chess would be pretty much unplayably drawish. But the win/draw line does seem to be uncomfortably close to a pawn, you don't need much comp to be able to hold with perfect play. Recently chess.com ran a series of games between top engines (alternating sides) with the f2 pawn removed from the start position, and Black won about half the games, with the others drawn except for one White win (out of about 30 games). With f2 gone the only compensation for the pawn is the first move, but this is fairly significant; f7 handicap is clearly lost for Black based both on analysis and on actual results of top engines.
Regarding the specific position in the thread, Komodo gives it a zero or near-zero score at low depths, which rises starting around depth 10 to a max of about half a pawn, which drops back to zero when reaching depth 25. Komodo MCTS gives a score around a quarter pawn after a second or so, which gradually declines asymptotically approaching zero after a while.

Do you believe that Komodo can become a stronger engine than Alpha Zero if you develop a combination of Neural Networks combined with either Alpha-beta Search or Pruning?

I wouldn't put it that way; rather I would say that I expect Komodo to become stronger than AlphaZero or Stockfish or Leela if we can combine Neural Networks and MCTS with human chess knowledge successfully. We don't yet have a working NN Komodo so it's only a hope at this point.

[Chessqueen]

For one, there is nothing wrong with a score of +1.53 and failing to win: the draw margin in Chess is about 1.5 Pawn, and +1.53 thus means the position offers about a 50-50 chance on win / draw.

I was very surprised to read this sentence. All of my tests in past years put the 50/50 win/draw line at about 3/4 of a pawn, not 1.5 pawns. Of course it depends a bit on which engine you are talking about and time limits, etc., but for Komodo (and Houdini too I think) 3/4 pawn is about right. A "clean" pawn up is evaluated as somewhat less than 1.0 by Komodo in the opening and somewhat more in the endgame, but on average 1.00 should be pretty close to accurate. The recent SF versions have very inflated scores (to say nothing of Lc0) which report a neutral pawn plus as more than 1.0, so if yu are basing this on recent SF that could be part of the problem, but even with SF I don't think the 50/50 line would be much over 1.0. So I'm asking what your source for the 1.5 pawn claim was; perhaps it was based on some very weak engine, or one that reports unrealistic scores for a pawn plus?

It was just a rule of thumb I used when I was still playing chess myself: being a clean Pawn ahead is usually not enough to win the game. (Of course I was just a patzer compared to you!) Most KRPKR are draw, with minors instead of Rooks it is even a dead draw. And in these cases the Pawn even is a passer, which should put it on the high end of the Pawn-value spectrum. Even KPK is often a draw, although without pieces advatages count much heavier.

My experience with weak engines (Fairy-Max ~2000 Elo) is that classical Pawn odds results in a 68% score (color-averaged). This doesn't seem to depend much on the level of play (as changed by varying the TC). I once saw a posting on the Rybka forum where someone had measured this with Rybka to be 70-72% (but I was not sure whether this was color averaged). Note that a 50-50 win:draw would be a 75% score, so a draw margin > 1 Pawn seems even to hold in the opening.

I don't know anythinhg about Komodo's evaluation scale; I don't have Komodo. More importantly, the 1.53 score was ascribed to LC0, so it really depends on how LC0 translates winning prospects into centi-Pawn. Which is just a matter of presentation, and has nothing to do with the quality of play: any monotonous transformation of the score would lead to exactly the same play.

Evaluating the given position as 0.00 seems a mistake: it means the engine will give up trying, even against a much weaker opponent. No matter how large a contempt you set. The contempt would only help to avoid immediate draws, but it would not help to distinguish between lines that keep the opponent under pressure, and those that blunder away one or two Pawns.

What would be Komodo's static evaluation of that position, btw?

High level (human or engine) chess is mostly about trying to win a pawn and then promote it. Grandmasters clearly believe that in general an extra pawn should win if there is no compensation or drawing factors; if this were not so chess would be pretty much unplayably drawish. But the win/draw line does seem to be uncomfortably close to a pawn, you don't need much comp to be able to hold with perfect play. Recently chess.com ran a series of games between top engines (alternating sides) with the f2 pawn removed from the start position, and Black won about half the games, with the others drawn except for one White win (out of about 30 games). With f2 gone the only compensation for the pawn is the first move, but this is fairly significant; f7 handicap is clearly lost for Black based both on analysis and on actual results of top engines.
Regarding the specific position in the thread, Komodo gives it a zero or near-zero score at low depths, which rises starting around depth 10 to a max of about half a pawn, which drops back to zero when reaching depth 25. Komodo MCTS gives a score around a quarter pawn after a second or so, which gradually declines asymptotically approaching zero after a while.

Do you believe that Komodo can become a stronger engine than Alpha Zero if you develop a combination of Neural Networks combined with either Alpha-beta Search or Pruning?

I wouldn't put it that way; rather I would say that I expect Komodo to become stronger than AlphaZero or Stockfish or Leela if we can combine Neural Networks and MCTS with human chess knowledge successfully. We don't yet have a working NN Komodo so it's only a hope at this point.

That is good News! That imply that your team is working on a NN Komodo version at this moment
I see that Pichy long time ago suggested the MCTS method that shogi program used to be implemented in Chess Engines. I have to give credit to Pichy for thinking about that idea. Anyway, what happened to Pichy, did he died?
viewtopic.php?f=3&t=46782&start=100

[Chessqueen]

For one, there is nothing wrong with a score of +1.53 and failing to win: the draw margin in Chess is about 1.5 Pawn, and +1.53 thus means the position offers about a 50-50 chance on win / draw.

I was very surprised to read this sentence. All of my tests in past years put the 50/50 win/draw line at about 3/4 of a pawn, not 1.5 pawns. Of course it depends a bit on which engine you are talking about and time limits, etc., but for Komodo (and Houdini too I think) 3/4 pawn is about right. A "clean" pawn up is evaluated as somewhat less than 1.0 by Komodo in the opening and somewhat more in the endgame, but on average 1.00 should be pretty close to accurate. The recent SF versions have very inflated scores (to say nothing of Lc0) which report a neutral pawn plus as more than 1.0, so if yu are basing this on recent SF that could be part of the problem, but even with SF I don't think the 50/50 line would be much over 1.0. So I'm asking what your source for the 1.5 pawn claim was; perhaps it was based on some very weak engine, or one that reports unrealistic scores for a pawn plus?

It was just a rule of thumb I used when I was still playing chess myself: being a clean Pawn ahead is usually not enough to win the game. (Of course I was just a patzer compared to you!) Most KRPKR are draw, with minors instead of Rooks it is even a dead draw. And in these cases the Pawn even is a passer, which should put it on the high end of the Pawn-value spectrum. Even KPK is often a draw, although without pieces advatages count much heavier.

My experience with weak engines (Fairy-Max ~2000 Elo) is that classical Pawn odds results in a 68% score (color-averaged). This doesn't seem to depend much on the level of play (as changed by varying the TC). I once saw a posting on the Rybka forum where someone had measured this with Rybka to be 70-72% (but I was not sure whether this was color averaged). Note that a 50-50 win:draw would be a 75% score, so a draw margin > 1 Pawn seems even to hold in the opening.

I don't know anythinhg about Komodo's evaluation scale; I don't have Komodo. More importantly, the 1.53 score was ascribed to LC0, so it really depends on how LC0 translates winning prospects into centi-Pawn. Which is just a matter of presentation, and has nothing to do with the quality of play: any monotonous transformation of the score would lead to exactly the same play.

Evaluating the given position as 0.00 seems a mistake: it means the engine will give up trying, even against a much weaker opponent. No matter how large a contempt you set. The contempt would only help to avoid immediate draws, but it would not help to distinguish between lines that keep the opponent under pressure, and those that blunder away one or two Pawns.

What would be Komodo's static evaluation of that position, btw?

High level (human or engine) chess is mostly about trying to win a pawn and then promote it. Grandmasters clearly believe that in general an extra pawn should win if there is no compensation or drawing factors; if this were not so chess would be pretty much unplayably drawish. But the win/draw line does seem to be uncomfortably close to a pawn, you don't need much comp to be able to hold with perfect play. Recently chess.com ran a series of games between top engines (alternating sides) with the f2 pawn removed from the start position, and Black won about half the games, with the others drawn except for one White win (out of about 30 games). With f2 gone the only compensation for the pawn is the first move, but this is fairly significant; f7 handicap is clearly lost for Black based both on analysis and on actual results of top engines.
Regarding the specific position in the thread, Komodo gives it a zero or near-zero score at low depths, which rises starting around depth 10 to a max of about half a pawn, which drops back to zero when reaching depth 25. Komodo MCTS gives a score around a quarter pawn after a second or so, which gradually declines asymptotically approaching zero after a while.

Do you believe that Komodo can become a stronger engine than Alpha Zero if you develop a combination of Neural Networks combined with either Alpha-beta Search or Pruning?

I wouldn't put it that way; rather I would say that I expect Komodo to become stronger than AlphaZero or Stockfish or Leela if we can combine Neural Networks and MCTS with human chess knowledge successfully. We don't yet have a working NN Komodo so it's only a hope at this point.

That is good News! That imply that your team is working on a NN Komodo version at this moment
I see that Pichy long time ago suggested the MCTS method that shogi program used to be implemented in Chess Engines. I have to give credit to Pichy for thinking about that idea. Anyway, what happened to Pichy, did he died?
viewtopic.php?f=3&t=46782&start=100

https://tcec.chessdom.com/
If Leela and Stockfish determine that it is a draw, why do they keep on playing on this position?
[D]6R1/8/8/2P5/8/KP3k2/P5pr/8 w - - 0 2