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

[lkaufman]

"The Evans is not in Connie's book
in playing the game over with NO QR odds ..Connie made every move exactly the same as in this QR odds except for this 7th move now
in a normal game she would play here 7..Bxc3

however in this game she played ..
7..dxc3"

I play 8Qb3, still normal theory with the rook on the board. I think 7...dxc3 is better than 7...Bxc3 with or without the rook, but especially with the rook, as in that case after 7...Bxc3 White can swap and use his queen's rook easily.

[Steve B]

I play 8Qb3, still normal theory with the rook on the board. I think 7...dxc3 is better than 7...Bxc3 with or without the rook, but especially with the rook, as in that case after 7...Bxc3 White can swap and use his queen's rook easily.

Connie desperately clings on to all of her ill-gotten gains with ...
8..Qf6

[d] r1b1k1nr/pppp1ppp/2n2q2/b7/2B1P3/1Qp2N2/P4PPP/1NB2RK1 w kq - 0 9

Material Advantage now an enormous...
+8.0 Regards
Steve

[Wayne Lowrance]

I like this dialogue. Very very good and constructive. I also will follow this experiment.
Wayne

[bob]

Wow, that is amazing! Can you even begin to explain how this is possible?

Sure. Don Beal wrote a paper on this many years ago. Unfortunately, his hardware at the time was very slow. His results today would have been _much_ better.

But here is the idea. Assume you have a PRNG that produces random numbers in the range 0 - .999, uniformly distributed. At the normal tips in the tree, you call this thing to get a random number and you back this up as if it were a real score.

Think about a position where you have exactly two legal moves. For the first move, you hang your queen, and after your opponent captures it, you get one of these random numbers, A big or small one? Uniformly distributed, so somewhere in the range.

Now for move 2, you save the queen, and now below that point you will have more legal moves to choose from. And each time you try a different move, you get a different random number. And the more of these you get, the greater the probability that you get a big one (a good score)

What this turns into is a "poor-man's mobility score" when you think about it. You will follow pathways in the tree that maximize your mobility (and hence your changes of getting some of those large random scores which will make you choose such moves) rather than pathways that minimize your mobility and hence your chances of getting a big random score. But the same token, you will tend to like moves that restrict your opponent's mobility because that gives him fewer chances to get a big score for himself.

As the tree gets deeper and bigger, this plays better and better. To the point that in the current version of crafty, compiled with -DSKILL so that you can set the skill from 1 (almost completely random) to 100 (no randomness at all, normal eval instead) and change the playing level. But the weakest player you can get is well in the 1800 Elo range. At skill level 1. And that is playing with just one CPU. Crank up 8 or 16 cores and that number is going up.

Pretty damned mind-boggling, although once you see the explanation, you think "of course, why didn't I see that." I felt the same way when Don published this in the ICCA journal several years ago...

It is possible that we could see IM-level play with fast enough hardware, or at least certainly within the next 5-10 years, with a one line evaluation function:

int Evaluate() {
return (random());
}



Not a bad result for so little effort.

What I have had to do to make the skill work is to knock the NPS _way_ back with a busy-loop so that the trees are far smaller to reduce this effect.

[Steve B]

I like this dialogue. Very very good and constructive. I also will follow this experiment.
Wayne

Glad you and others seem to find this interesting
the GM usually posts his move by now
about 8 hrs to go for the 24 hour TC rule(as i write)

make no mistake..
i am not above claiming a win by TIME FORFEIT


J/K Regards
Steve

[lkaufman]

I play 8Qb3, still normal theory with the rook on the board. I think 7...dxc3 is better than 7...Bxc3 with or without the rook, but especially with the rook, as in that case after 7...Bxc3 White can swap and use his queen's rook easily.

Connie desperately clings on to all of her ill-gotten gains with ...
8..Qf6

[d] r1b1k1nr/pppp1ppp/2n2q2/b7/2B1P3/1Qp2N2/P4PPP/1NB2RK1 w kq - 0 9

Material Advantage now an enormous...
+8.0 Regards
Steve

I play 9.e5. An alternative was 9.Bg5. Both moves lead to an unclear game with the rook on the board. Although I think my chances now are better than they would have been if I played some boring opening, I will sorely miss the queen's rook once I've played Nxc3 and developed by queen's bishop.But I'd miss it in almost any opening!

[lkaufman]

"Sure. Don Beal wrote a paper on this many years ago. Unfortunately, his hardware at the time was very slow. His results today would have been _much_ better.

But here is the idea. Assume you have a PRNG that produces random numbers in the range 0 - .999, uniformly distributed. At the normal tips in the tree, you call this thing to get a random number and you back this up as if it were a real score.

Think about a position where you have exactly two legal moves. For the first move, you hang your queen, and after your opponent captures it, you get one of these random numbers, A big or small one? Uniformly distributed, so somewhere in the range.

Now for move 2, you save the queen, and now below that point you will have more legal moves to choose from. And each time you try a different move, you get a different random number. And the more of these you get, the greater the probability that you get a big one (a good score)

What this turns into is a "poor-man's mobility score" when you think about it. You will follow pathways in the tree that maximize your mobility (and hence your changes of getting some of those large random scores which will make you choose such moves) rather than pathways that minimize your mobility and hence your chances of getting a big random score. But the same token, you will tend to like moves that restrict your opponent's mobility because that gives him fewer chances to get a big score for himself.

As the tree gets deeper and bigger, this plays better and better. To the point that in the current version of crafty, compiled with -DSKILL so that you can set the skill from 1 (almost completely random) to 100 (no randomness at all, normal eval instead) and change the playing level. But the weakest player you can get is well in the 1800 Elo range. At skill level 1. And that is playing with just one CPU. Crank up 8 or 16 cores and that number is going up."

Sure, this makes sense. In fact I now remember the article about this by Don Beal so long ago. Still, 1800 is surprising.
I wonder what would be the rating (on similar hardware and time limit) for an eval that did only material (1-3-3-5-9), and for an eval that considered only the number of legal moves? It's not obvious how these three evals (including random) would rank. Has anyone done these tests?

[Steve B]

I play 9.e5. An alternative was 9.Bg5. Both moves lead to an unclear game with the rook on the board. Although I think my chances now are better than they would have been if I played some boring opening, I will sorely miss the queen's rook once I've played Nxc3 and developed by queen's bishop.But I'd miss it in almost any opening!

Well one thing is for certain
there is nothing boring about this game

Connie walks a very thin line between the thrill of victory and the agony of defeat with..

9..Qf5

[d] r1b1k1nr/pppp1ppp/2n5/b3Pq2/2B5/1Qp2N2/P4PPP/1NB2RK1 w kq - 0 10

Delicate Balancing Act Regards
Steve

[bob]

"Sure. Don Beal wrote a paper on this many years ago. Unfortunately, his hardware at the time was very slow. His results today would have been _much_ better.

But here is the idea. Assume you have a PRNG that produces random numbers in the range 0 - .999, uniformly distributed. At the normal tips in the tree, you call this thing to get a random number and you back this up as if it were a real score.

Think about a position where you have exactly two legal moves. For the first move, you hang your queen, and after your opponent captures it, you get one of these random numbers, A big or small one? Uniformly distributed, so somewhere in the range.

Now for move 2, you save the queen, and now below that point you will have more legal moves to choose from. And each time you try a different move, you get a different random number. And the more of these you get, the greater the probability that you get a big one (a good score)

What this turns into is a "poor-man's mobility score" when you think about it. You will follow pathways in the tree that maximize your mobility (and hence your changes of getting some of those large random scores which will make you choose such moves) rather than pathways that minimize your mobility and hence your chances of getting a big random score. But the same token, you will tend to like moves that restrict your opponent's mobility because that gives him fewer chances to get a big score for himself.

As the tree gets deeper and bigger, this plays better and better. To the point that in the current version of crafty, compiled with -DSKILL so that you can set the skill from 1 (almost completely random) to 100 (no randomness at all, normal eval instead) and change the playing level. But the weakest player you can get is well in the 1800 Elo range. At skill level 1. And that is playing with just one CPU. Crank up 8 or 16 cores and that number is going up."

Sure, this makes sense. In fact I now remember the article about this by Don Beal so long ago. Still, 1800 is surprising.
I wonder what would be the rating (on similar hardware and time limit) for an eval that did only material (1-3-3-5-9), and for an eval that considered only the number of legal moves? It's not obvious how these three evals (including random) would rank. Has anyone done these tests?

No, but it is an idea I will test as I find time on a cluster that is acting up due to significant A/C issues. It may well be that scaling down the randomness would help, and I have not tried that.

I'll try pure material, number of moves, and pure random, and perhaps a scaled down random (say 0, 1, 2, 3, 4, 5) which will eliminate some of that "pseudo-mobility" stuff...

[BubbaTough]

No, but it is an idea I will test as I find time on a cluster that is acting up due to significant A/C issues. It may well be that scaling down the randomness would help, and I have not tried that.

I'll try pure material, number of moves, and pure random, and perhaps a scaled down random (say 0, 1, 2, 3, 4, 5) which will eliminate some of that "pseudo-mobility" stuff...

Are you still detecting mate and stalemate normally?

-Sam