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

[lkaufman]

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

I play 10.Nxc3. Aside from winning back a pawn, at least Black has to "pay" the bishop pair if she wants to exchange. I must avoid truly even trades like the plague. I think that 9...Qg6 is considered slightly more accurate in normal chess, but Connie rightly avoided 9...Nxe5 10Re1 d6 11.Qa4ch, though even in that case Connie would be up the Exchange and a boatload of pawns. This would only take a 3 or 4 ply search to see, so I was pretty sure Connie would not take this pawn. So far Connie is not playing like a "B" player, but like a very greedy master. Unless she starts making mistakes soon, I'm in big trouble.

[PK]

As for experiments, I have tried the following (starting with open source program Sungorus):

- eval function = material + pst + large random component (currently 500cp)
- no transposition table (which would flatten results)
- at root we do n searches and a move gets the average score

so far n=10 beats n=20, and n=40 beats both of them. the highest I have tried is 80 - and it plays something that looks like chess, but its strength is probably about 900 Elo (compared to weaker personalities of as-yet-unreleased Glass 1.6, so not to reliable). If I ever manage to get to 1800, it might be worth releasing as a funny mix of randomness and alpha-beta.

[Steve B]

I play 10.Nxc3. Aside from winning back a pawn, at least Black has to "pay" the bishop pair if she wants to exchange. I must avoid truly even trades like the plague. I think that 9...Qg6 is considered slightly more accurate in normal chess, but Connie rightly avoided 9...Nxe5 10Re1 d6 11.Qa4ch, though even in that case Connie would be up the Exchange and a boatload of pawns. This would only take a 3 or 4 ply
search to see, so I was pretty sure Connie would not take this pawn. So far Connie is not playing like a "B" player, but like a very greedy master. Unless she starts making
mistakes soon, I'm in big trouble.

The Novag Constellation 3.6 Mhz prepares for King-Side Castling(i hope)
with..
10..Nge7

[d] r1b1k2r/ppppnppp/2n5/b3Pq2/2B5/1QN2N2/P4PPP/2B2RK1 w kq - 0 11

As we have Now reached the 10th move ..i include the current PGN:

REMOVE WHITE QR
[Event "CCC Forum Challenge "]
[Date "2010.7.3"]
[White "GM Kaufman"]
[Black "Novag Constellation 3.6"]
[Time Control "3 Min./Avg."]
[Result "*"]

1.e4 Nc6 2.Nf3 e5 3.Bc4 Bc5 4.b4 Bxb4 5.c3 Ba5 6.d4 exd4 7.O-O dxc3
8.Qb3 Qf6 9.e5 Qf5 10.Nxc3 Nge7 *

Sunday Morning Brunch With Connie Regards
Steve

[gerold]

I play 10.Nxc3. Aside from winning back a pawn, at least Black has to "pay" the bishop pair if she wants to exchange. I must avoid truly even trades like the plague. I think that 9...Qg6 is considered slightly more accurate in normal chess, but Connie rightly avoided 9...Nxe5 10Re1 d6 11.Qa4ch, though even in that case Connie would be up the Exchange and a boatload of pawns. This would only take a 3 or 4 ply
search to see, so I was pretty sure Connie would not take this pawn. So far Connie is not playing like a "B" player, but like a very greedy master. Unless she starts making
mistakes soon, I'm in big trouble.

The Novag Constellation 3.6 Mhz prepares for King-Side Castling(i hope)
with..
10..Nge7

[d] r1b1k2r/ppppnppp/2n5/b3Pq2/2B5/1QN2N2/P4PPP/2B2RK1 w kq - 0 11

As we have Now reached the 10th move ..i include the current PGN:

REMOVE WHITE QR
[Event "CCC Forum Challenge "]
[Date "2010.7.3"]
[White "GM Kaufman"]
[Black "Novag Constellation 3.6"]
[Time Control "3 Min./Avg."]
[Result "*"]

1.e4 Nc6 2.Nf3 e5 3.Bc4 Bc5 4.b4 Bxb4 5.c3 Ba5 6.d4 exd4 7.O-O dxc3
8.Qb3 Qf6 9.e5 Qf5 10.Nxc3 Nge7 *

Sunday Morning Brunch With Connie Regards
Steve

Thanks for the game. Nice.

I gave Connie a full point for the last move.
Now Black can castle and run away with the game

Best,
Gerold.

[bob]

n

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

Yes. That's done in the search, which I did not change at all. Just the endpoint evaluation is the issue, but the randomness is killing the ability to get the Elo down to the desired 800 or less. The search is greatly restricted however, because LMR, null-move and check extensions are gone by the time you get down to skill 1.

[BubbaTough]

n

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

Yes. That's done in the search, which I did not change at all. Just the endpoint evaluation is the issue, but the randomness is killing the ability to get the Elo down to the desired 800 or less. The search is greatly restricted however, because LMR, null-move and check extensions are gone by the time you get down to skill 1.

Ahh, I see. So its an Elo 800 that consistently sees mate in 5 for itself or its opponent (this is not meant to be a criticism, I have heard a lot of complements about Crafty's skill level mode and am curious how pervasive your code changes are to support it). Do you also have razoring or material based pruning that does not invoke eval?

-Sam

[bob]

n

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

Yes. That's done in the search, which I did not change at all. Just the endpoint evaluation is the issue, but the randomness is killing the ability to get the Elo down to the desired 800 or less. The search is greatly restricted however, because LMR, null-move and check extensions are gone by the time you get down to skill 1.

Ahh, I see. So its an Elo 800 that consistently sees mate in 5 for itself or its opponent (this is not meant to be a criticism, I have heard a lot of complements about Crafty's skill level mode and am curious how pervasive your code changes are to support it). Do you also have razoring or material based pruning that does not invoke eval?

-Sam

There are some non-eval things, but skill pares all the selective search stuff back. Cuts the depth by 60-70% depending on position. It is tough to find a mate in 5 with no extensions and no pruning, which limits the depth to 6-7 plies in 10 sec games...

[lkaufman]

The Novag Constellation 3.6 Mhz prepares for King-Side Castling(i hope)
with..
10..Nge7

[d] r1b1k2r/ppppnppp/2n5/b3Pq2/2B5/1QN2N2/P4PPP/2B2RK1 w kq - 0 11

I play 11.Ba3, which would normally prepare to develop the queen's rook to d1. Here it at least the bishop will pin the knight if Black castles, and will discourage the freeing move ...d6. Now I wonder if Black will castle or trade on c3 first. So far no clear errors by Black.

[Albert Silver]

The Novag Constellation 3.6 Mhz prepares for King-Side Castling(i hope)
with..
10..Nge7

[d] r1b1k2r/ppppnppp/2n5/b3Pq2/2B5/1QN2N2/P4PPP/2B2RK1 w kq - 0 11

I play 11.Ba3, which would normally prepare to develop the queen's rook to d1. Here it at least the bishop will pin the knight if Black castles, and will discourage the freeing move ...d6. Now I wonder if Black will castle or trade on c3 first. So far no clear errors by Black.

I thought you might go for Ne2 to avoid the possible piece swap and try to take advantage of Black's queen wandering around.

[Don]

Good case is the LMR stuff. I looked briefly at the Stockfish stuff of trying to increase the reduction as you complete more and more moves. But here's a question. Given two sets of moves, MA(i) and MB(i) where both MA and MB have the same number of moves, but in MA the first couple are good and the rest are lemons, while in MB all but the last couple are good. What is the justification for reducing the middle move in MA the same as the middle move in MB, when the middle move in MA is bad while the middle move in MB is good? No justification at all. Just force-fitting a linear value to a polynomial data set, and finding the best liner approximation you can get, even though it is bad in many cases.

I do not like the idea of reducing (or extending) something just because it comes before or after the Nth move in the list. Surely there must be a better reason. I certainly do not do that kind of nonsense as a human. I believe we can do this stuff pretty well (fit the wrong type of curve, but at least optimize the poor fit so that we arrive at the best "poor fit" we can get.)

I think the principle of waiting until the Nth move and then reducing has powerful justification. It turns out that modern chess programs with good move ordering have the characteristic that the first move is by far more likely to the best move than the second, and the second is far more likely than the 3rd and so on. By the time you have gotten to the Nth move, the probability that you have found a good (if not the best) move is extremely high.

But it's not like we just throw out the rest, we evaluate them with a reduced depth search as a "test" to see if they should be search full depth.

So we are actually inspecting every single move pretty thoroughly and leaning on the fact that the best move is almost always one of the first moves considered.

Another way to view this is to speculate how you might improve on this decision without considering a moves position in the move list. The only reasonable thing I can imagine is evaluating each move with a reduced depth search. That is our very best way of evaluating moves. But the hash table move and often the killers are just as good or better than doing that.

Another criteria is dynamic considerations - but we already do that - captures and checks are not reduced either. So what is left?

I think the position in the move list is an extremely powerful notion and it's combined with a thorough evaluation on top of that (the reduced depth search) so it's very difficult to improve on this.

You are troubled by the arbitrariness of N and I agree. So you could consider making N variable depending on many other factors such as depth, alpha/beta, score of position, stage of game, type of position and characteristics of moves already tried.

In fact we do a little of that. For LMR purposes we don't count moves that we would not normally reduce anyway. So we require N "quiet" moves before reducing. That is a small but measurable improvement for us and makes N slightly less arbitrary.