GM Kaufman v.Novag Constellation 3.6 QR Odds (Moves 1-40)

[Steve B]

I play the expected 31.h5. Unfortunately there is no chance that Connie might be tempted to take the pawn, as it loses the queen to a discovered check. I'm leaving in the morning for the US Open, and while I should still have internet access I might not be quite as prompt about making my moves as I have been so far. I'll try to move once a day if possible.

Connie always eager to harass the White K plays ..
31..Qd6+

[d] 4r1k1/p1pr1p1p/1p1q2p1/3n2BP/2Bp2R1/1Q6/P4P1K/8 w - - 0 32

Always Check ..It could be mate regards
Steve

[Terry McCracken]

I play the expected 31.h5. Unfortunately there is no chance that Connie might be tempted to take the pawn, as it loses the queen to a discovered check. I'm leaving in the morning for the US Open, and while I should still have internet access I might not be quite as prompt about making my moves as I have been so far. I'll try to move once a day if possible.

Connie always eager to harass the White K plays ..
31..Qd6+

[d] 4r1k1/p1pr1p1p/1p1q2p1/3n2BP/2Bp2R1/1Q6/P4P1K/8 w - - 0 32

Always Check ..It could be mate regards
Steve

32. Kg2..Qc6 33. Kh2..Re8 34. hg..hg ........

[bob]

That will be an interesting test. Computers are, in general, horrible odds players. You can't keep exchanging a pair of pieces to win a pawn. You run out of pieces before you catch back up in material, leaving yourself lost.

thats exactly correct
i have played 6 games now against the same 2700 PC engine at R odds
the lower rated dedicated's(all 1700 range) have won 2 and drawn 4
i only saved 3 of the games(1 win- 2 draws)
i initially tried some games(about 6 or so) at "time odds" and "minor piece" odds but the dedicated's could not draw or win even one game
in each"R odds" game the Engine has usually won back material quickly enough but wound up either lost or hoplessly drawn due to the reduced material

thats what i suspect will happen in the Movei game coming up after this game (but it will be interesting to play against a different engine)

Steve

A human will think "OK, I am already objectively lost, so winning back a pawn is pointless. My goal right now is to not trade material and try to build up an attack to break thru, otherwise the endgame is lost completely." A program is going to win anything it can. If it can trade queens to wreck your pawn structure, it will, even though your rook advantage will nullify (and then some) any pawn structure weakness. But hey, it gave you 3 isolated pawns, even if it did give up all attacking chances.

This has been a hard idea to get across. When behind, trade pawns but not pieces, when ahead, trade pieces but not pawns. When we say "not pieces" we _really_ mean it, but to the programs it is just a weak suggestion, at best.

[Milos]

That will be an interesting test. Computers are, in general, horrible odds players. You can't keep exchanging a pair of pieces to win a pawn. You run out of pieces before you catch back up in material, leaving yourself lost.

Which raises an idea. Why not include this in material evaluation. There is already a second order effect in the phase calculation, but this is not sufficient.
When you exchange pieces you get into later phase, meaning that remaining peaces and also odd you have is worth more. However, the value is smaller than 100cp.
So the idea would be to scale material score with total sum of pieces you have.
One example formula can be:
new_material_score=material_score*material_score*(starting_total_material-material_score)/((material_score-1)*your_current_material + starting_total_material-material_score)

So, taking base material values 3, 3, 6, 9, if you have rook odd at opening, opponent engine would value:
new_material_score=6*6*21/(5*21+21)=6
After exchanging knights for example it would be:
new_material_score=6*6*21/(5*18+21)=6.81
Combined with later phase this would be enough to prevent exchanges for a pawn when you are rook down.
However, in case of knight odd:
new_material_score=3*3*24/(2*24+24)=3
After exchanging bishops for example it would be:
new_material_score=3*3*24/(2*21+24)=3.27
So it's worth exchange to win a pawn.
However, if engine has rook and knight while his opponent has rook and two knights, exchanging knights brings:
before exchange: new_material_value=3*3*24/(2*9+24)=5.14
after exchange: new_material_value=3*3*24/(2*6+24)=6
So, after phase adjustment pawn gain is not sufficient to compensate for the exchange.

[Don]

That will be an interesting test. Computers are, in general, horrible odds players. You can't keep exchanging a pair of pieces to win a pawn. You run out of pieces before you catch back up in material, leaving yourself lost.

thats exactly correct
i have played 6 games now against the same 2700 PC engine at R odds
the lower rated dedicated's(all 1700 range) have won 2 and drawn 4
i only saved 3 of the games(1 win- 2 draws)
i initially tried some games(about 6 or so) at "time odds" and "minor piece" odds but the dedicated's could not draw or win even one game
in each"R odds" game the Engine has usually won back material quickly enough but wound up either lost or hoplessly drawn due to the reduced material

thats what i suspect will happen in the Movei game coming up after this game (but it will be interesting to play against a different engine)

Steve

A human will think "OK, I am already objectively lost, so winning back a pawn is pointless. My goal right now is to not trade material and try to build up an attack to break thru, otherwise the endgame is lost completely." A program is going to win anything it can. If it can trade queens to wreck your pawn structure, it will, even though your rook advantage will nullify (and then some) any pawn structure weakness. But hey, it gave you 3 isolated pawns, even if it did give up all attacking chances.

This has been a hard idea to get across. When behind, trade pawns but not pieces, when ahead, trade pieces but not pawns. When we say "not pieces" we _really_ mean it, but to the programs it is just a weak suggestion, at best.

This same basic issue comes up in computer GO with the new and modern Monte Carlo Tree Search. Monte Carlo of course plays thousands of very fast games as an estimate of the value of the position. But in GO you can score wins and losses, or you can actually use the "degree" of the win by tallying up territory at the end of the game. In chess you either checkmate or you don't. But in Go you could win big or win small.

It turns out that to get the best results you need to simply tally up wins and losses. A big win is the same as a small win. When you score the Monte Carlo playouts in this way you get significantly better results.

There is one nasty side effect of doing this however is that if you have a game that is basically lost, the program see's the position as hopeless and will not play "normal" fighting moves. You could be losing by just 1 stone but the program doesn't care if it's 1 stone or 100 and it will go down without a fight.

From the program's point of view there are no good or bad moves because they all lose.

If a Monte Carlo Chess program were playing this game in Larry's place, it would see certain loss and would play like Terry McCracken, noticing all the over the board issues while being completely oblivious to the more important "opponent modeling" issues which define what heavy odds games are all about. It would basically say, "I'm dead lost" and it would roll over and die - not realizing what the real game is.

A thought experiment is to play a long chess match where points are scored based on the move number you checkmate your opponent on - the sooner you checkmate the more points you score. It's suddenly a different game, and the rules of chess are a subset of this game. A computer would have to be re-programmed in order to play this game well.

[bob]

That will be an interesting test. Computers are, in general, horrible odds players. You can't keep exchanging a pair of pieces to win a pawn. You run out of pieces before you catch back up in material, leaving yourself lost.

Which raises an idea. Why not include this in material evaluation. There is already a second order effect in the phase calculation, but this is not sufficient.
When you exchange pieces you get into later phase, meaning that remaining peaces and also odd you have is worth more. However, the value is smaller than 100cp.
So the idea would be to scale material score with total sum of pieces you have.
One example formula can be:
new_material_score=material_score*material_score*(starting_total_material-material_score)/((material_score-1)*your_current_material + starting_total_material-material_score)

So, taking base material values 3, 3, 6, 9, if you have rook odd at opening, opponent engine would value:
new_material_score=6*6*21/(5*21+21)=6
After exchanging knights for example it would be:
new_material_score=6*6*21/(5*18+21)=6.81
Combined with later phase this would be enough to prevent exchanges for a pawn when you are rook down.
However, in case of knight odd:
new_material_score=3*3*24/(2*24+24)=3
After exchanging bishops for example it would be:
new_material_score=3*3*24/(2*21+24)=3.27
So it's worth exchange to win a pawn.
However, if engine has rook and knight while his opponent has rook and two knights, exchanging knights brings:
before exchange: new_material_value=3*3*24/(2*9+24)=5.14
after exchange: new_material_value=3*3*24/(2*6+24)=6
So, after phase adjustment pawn gain is not sufficient to compensate for the exchange.

Most do this already. We certainly do. That is, given a position where white is behind by some amount of material, the score is worse given the same material balance, but with fewer pieces total. And the score is better given the same material balance but with fewer pawns.

But in an odds game, this "penalty" needs to be really large, because winning a pawn is not enough when down a rook. It is good if you can just take the thing, of course. But if you have to trade any pieces to pick up the pawn, it is likely a bad idea. That's where the problem comes in, since you start off behind, rather than slowly falling behind (or ahead). Normally winning a pawn is a pretty good idea, even if you have to trade 1/2 your pieces. But in this odds stuff, that changes.

Certainly one could modify the evaluation to do better in odds games, by making these penalties much larger, since you would never start off "even" anyway. But then you would likely have problems in normal games.

[Milos]

Certainly one could modify the evaluation to do better in odds games, by making these penalties much larger, since you would never start off "even" anyway. But then you would likely have problems in normal games.

Not necessarily. If you more carefully examine the function I gave up there, it stimulates exchanges if you are better in material (actually even if you are worse up to 1 pawn). The amount of "stimulation" depends quadratically on material score.
If you are a rook down there will be really strong discouragement to make exchanges, but if you are much less down (like a bishop pair for example) nothing would change. On the other hand any kind of positive material score will strongly encourage exchanges.
I'll test the upper function and see how it goes. My feeling (which ofc can be completely off) tells me it could help a bit even in normal games without odds.

[bob]

Certainly one could modify the evaluation to do better in odds games, by making these penalties much larger, since you would never start off "even" anyway. But then you would likely have problems in normal games.

Not necessarily. If you more carefully examine the function I gave up there, it stimulates exchanges if you are better in material (actually even if you are worse up to 1 pawn). The amount of "stimulation" depends quadratically on material score.
If you are a rook down there will be really strong discouragement to make exchanges, but if you are much less down (like a bishop pair for example) nothing would change. On the other hand any kind of positive material score will strongly encourage exchanges.
I'll test the upper function and see how it goes. My feeling (which ofc can be completely off) tells me it could help a bit even in normal games without odds.

That "quadratically" is what I am worried about it. You don't want to get ahead and then start wrecking your position to try to trade. That's a common flaw when this term gets out of whack. By the same token, you don't want to wreck your position to avoid trading either...

You sort of need to "ease into" those scores. But once you are a rook down or so, the game is lost (assuming the search doesn't spot a saving tactic as you are going down a rook).

While it might well be possible to come up with a "one-size-fits-all" approach, as a human I simply don't play like that.

[lkaufman]

Connie always eager to harass the White K plays ..
31..Qd6+

[d] 4r1k1/p1pr1p1p/1p1q2p1/3n2BP/2Bp2R1/1Q6/P4P1K/8 w - - 0 32

Always Check ..It could be mate regards
Steve

I play the natural 32.Kg2 by process of elimination; no point in allowing ...Re1 with check, and Kh3 or f4 look ridiculous.

[Steve B]

I play the natural 32.Kg2 by process of elimination; no point in allowing ...Re1 with check, and Kh3 or f4 look ridiculous.

Connie goes in for a bit of subtle maneuvering with ..

32..Qc6

[d] 4r1k1/p1pr1p1p/1pq3p1/3n2BP/2Bp2R1/1Q6/P4PK1/8 w - - 0 32

Finessing Regards
Steve