GM says Rybka & Fritz weaker than best GMs in classical

[George Tsavdaris]

Humans have _far_ superior positional understanding. By a factor of at least 10x and probably more than 100X better than any program around.

I agree about the 10X, but 100X? No way.
If you feed programs with a positional test suite you will notice that todays programs find a good percentage of the correct moves, much more than the past.

I would agree that perhaps it is for the wrong reasons, for example very deep tactics and not because they can plan for the far future as humans do when playing a positional move, but wrong or correct reasons, fact is they play many of the positional moves.

Also if you say that positional understanding of humans is about 10X to 100X better relative to computers, then i think you should emphasize the fact that computers are 500X to 1000X times better in all kind of tactics.

Been that way for a long time, and nothing has really changed, except we are beginning to see the tactical frailness of human chess over a 3-4-5 hour game is the limiting factor that is making programs superior. But that is the _only_ reason...

I agree this is the only reason that top humans lose to computers these days, but i disagree that things has not changed for computers about playing positional moves.

Rybka for example, plays many amazing positional moves, mostly exchange type sacrifices and some times not because a 14 ply tactic gives back the material, but long term ones.
I remember a game where Rybka gave its Rook for a Bishop(not any Pawn) only to squeeze only to win 25 moves later the material. That was a true positional sacrifice. There are many such examples and not only with Rybka.

For example(i think this is a bad example):
Rybka here played the amazing 12...b3!! that wins.
[D]rnbqk2r/p3ppb1/3p3p/P1pP2pn/1pP1P3/5NB1/RP1N1PPP/3QKB1R b Kkq - 0 12

After the almost forced for white 13.Nxb3 Na6 14.Qb1 f5! 15.exf5 0-0 16.Bd3 Nb4 black is controlling the board.
Perhaps this 12...b3!! is a bad example of positional play and it's just a deep tactic, but in fact what we say positional play is just incredibly deep tactics!
Humans play some positional moves because they know(or they believe from experience) many moves later they will have the advantage.

[Uri]

How well do computers understand both the position and the game?

I saw the win faster than a computer. I saw it on the spot!

[d]r1bq1r1k/1pp3bp/8/p3pp2/2P3pn/P1NPP3/1BQ2PPP/R3RBK1 w - - 0 20

Zoltan Ribli vs Eric Lobron

Shredder 10 doesn't see the sixth-rank rook lift but i'm sure a more aggressive and more powerful program like Deep Hiarcs 12, Rybka 2.3.2a mp or Deep Shredder XP will see it.

This is a classic example of attack on the king which Aleksey Bartashnikov explains quite well in the "Basic principles of chess strategy volume 2". Daniel King also explains this in "Power Play 2: Attacking the king".

[George Tsavdaris]

How well do computers understand both the position and the game?

I saw the win faster than a computer. I saw it on the spot!

[d]r1bq1r1k/1pp3bp/8/p3pp2/2P3pn/P1NPP3/1BQ2PPP/R3RBK1 w - - 0 20

My current computer is slow relative to the QUAD computers etc. It's about 12 times slower than a QX9770.
Yet Bright 0.3a for example finds Ra6 in 6 minutes. That means it would find Ra6 in about 30 seconds in a top QUAD.

Analysis by bright-0.3a:

20...Rf8-e8 21.Bf1-e2 Bc8-e6 22.g2-g3 Nh4-f3+ 23.Be2xf3 g4xf3
= (0.09) Depth: 6/19 00:00:01 35kN
20...Rf8-e8 21.Bf1-e2 Bc8-e6 22.Qc2-b3 Qd8-b8 23.d3-d4 e5xd4 24.e3xd4
= (0.09) Depth: 7/19 00:00:01 79kN
20...Rf8-e8 21.Nc3-e2 Bc8-d7 22.d3-d4 e5xd4 23.Ne2xd4 Qd8-e7 24.c4-c5 Qe7-e5
+/= (0.27) Depth: 8/23 00:00:01 236kN
20...Rf8-e8 21.Nc3-e2 Qd8-g5 22.c4-c5 b7-b6 23.d3-d4 Bc8-b7 24.Ne2-g3 f5-f4
= (0.17) Depth: 9/25 00:00:02 739kN
20...Rf8-e8 21.Nc3-e2 Qd8-g5 22.c4-c5 b7-b6 23.d3-d4 b6xc5 24.d4xe5 Ra8-b8 25.Ne2-g3 Bg7xe5 26.Bb2xe5+ Re8xe5
+/= (0.27) Depth: 10/28 00:00:02 1393kN
20...Rf8-e8 21.d3-d4 e5xd4 22.e3xd4 Bc8-d7 23.Nc3-d5 c7-c6 24.Re1xe8+ Bd7xe8 25.Nd5-e3 Qd8-f6 26.Bf1-d3 Be8-d7
+/= (0.27) Depth: 11/31 00:00:07 4171kN
20...Rf8-e8 21.d3-d4 e5xd4 22.e3xd4 Bc8-d7 23.Nc3-d5 Re8xe1 24.Rd1xe1 c7-c6 25.Nd5-f4 Qd8-c7 26.Nf4-h5 Ra8-e8 27.Re1xe8+ Bd7xe8 28.Nh5xg7 Qc7xg7
+/= (0.40) Depth: 12/34 00:00:12 7053kN
20...Rf8-e8 21.d3-d4 e5xd4 22.Nc3-b5 c7-c5 23.e3xd4 Re8xe1 24.Rd1xe1 c5xd4 25.Nb5xd4 Bc8-d7 26.Nd4-e6 Bd7xe6 27.Re1xe6 Bg7-d4
+/= (0.61) Depth: 13/38 00:00:36 20453kN
20...b7-b6 21.d3-d4 e5xd4 22.Nc3-d5 Bc8-b7 23.Bb2xd4 Bg7xd4 24.Rd1xd4 Bb7xd5 25.Rd4xd5 Qd8-f6 26.Re1-d1 Rf8-d8 27.Rd5-d7 Rd8xd7 28.Rd1xd7 Qf6-e5
+/= (0.45) Depth: 13/38 00:00:48 26108kN
20...b7-b6 21.d3-d4 e5xd4 22.e3xd4 Bc8-d7 23.d4-d5 Rf8-e8 24.Re1xe8+ Bd7xe8 25.Nc3-e2 Qd8-g5 26.Bb2xg7+ Qg5xg7 27.Ne2-d4 Qg7-e5 28.Bf1-e2 Be8-d7
+/= (0.46) Depth: 14/40 00:01:30 47901kN
20...b7-b6 21.d3-d4 e5xd4 22.e3xd4 Bc8-d7 23.Nc3-d5 Bd7-c6 24.Nd5-f4 Bc6-e4 25.Bf1-d3 Be4xd3 26.Rd1xd3 Qd8-d6 27.Nf4-e6 Ra8-e8 28.d4-d5 Rf8-f7 29.Bb2xg7+ Rf7xg7
+/= (0.53) Depth: 15/45 00:02:55 104mN
20...Ra8-a6 21.Nc3-e2 Bc8-d7 22.d3-d4 Qd8-e8 23.Qc2-b1 Bd7-a4 24.Rd1-d2 Qe8-d7 25.Ne2-c3 Ba4-c6 26.d4-d5 Bc6-a4 27.Nc3xa4 Qd7xa4
+/= (0.28) Depth: 15/50 00:06:02 226mN


Also even if no computer would be able to find the move, that would not mean anything about the GM-computers controversy.....

[Terry McCracken]

How well do computers understand both the position and the game?

I saw the win faster than a computer. I saw it on the spot!

[d]r1bq1r1k/1pp3bp/8/p3pp2/2P3pn/P1NPP3/1BQ2PPP/R3RBK1 w - - 0 20

Zoltan Ribli vs Eric Lobron

Shredder 10 doesn't see the sixth-rank rook lift but i'm sure a more aggressive and more powerful program like Deep Hiarcs 12, Rybka 2.3.2a mp or Deep Shredder XP will see it.

This is a classic example of attack on the king which Aleksey Bartashnikov explains quite well in the "Basic principles of chess strategy volume 2". Daniel King also explains this in "Power Play 2: Attacking the king".

Hiracs 11.2 sees it, Toga 1.4.1 SE can't see it, at least on modest hardware, a 1.8 ghz chip.

There are still many positions the machines seem to be baffled by.

This may not be a good test. However, I saw the solution straight away.

Computers seem to have some trouble here.

[Michael Sherwin]

How well do computers understand both the position and the game?

I saw the win faster than a computer. I saw it on the spot!

[d]r1bq1r1k/1pp3bp/8/p3pp2/2P3pn/P1NPP3/1BQ2PPP/R3RBK1 w - - 0 20



Zoltan Ribli vs Eric Lobron

If instead of Ne2 as in the game, what if Nd5 were played with the idea of Nf6 blocking the path of black's QR. Just a quick look, but looks winning for white.

[bob]

I'll say it again... can you tell the difference between positional understanding and tactical skill? I can. Every computer makes horrible positional moves, but they cover up these mistakes by exploiting every minor tactical mistake the human makes...

[bob]

Humans have _far_ superior positional understanding. By a factor of at least 10x and probably more than 100X better than any program around.

That is of course 5...50 times intentionally exaggerated !
Of what use is positional understanding if a sequence of sub-optimal tactics throws the game out of the window ?

Matthias.

Again experience over guesswork here. Under the right circumstances, such as the old correspondence games Mike Valvo used to play against Deep Thought, you can see just how wide the gap actually is. Humans make _very_ few tactical mistakes in correspondence. And there their positional judgement is the critical component...

As I said, computer tactical accuracy is what is winning games today, not positional judgement where they are still sorely lacking. However, at today's speeds, tactical accuracy is obviously quite enough to more than compensate for the ugly positional mistakes (most of the time, although humans still win games).

[bob]

In the July 2008 edition of Chess Life (page 44), GM Lev Alburt
claims "Playing under classical conditions (40 moves in 2 1/2 hours),
today's best grandmasters are still favorites vs. Rybka and Fritz".

I was under the opposite impression, but I'm not a GM.

Sounds like a challenge for some program to me.

Just because someone is a GM, that does not mean he has a clue what he is talking about.
We want to see proofs, not baseless claims !
The top engines are proving their superiority more and more convincingly.

Matthias.

If you understood his point, the discussion would already be over.

Humans have _far_ superior positional understanding. By a factor of at least 10x and probably more than 100X better than any program around. however, the programs have far superior tactical skill, when the effort is spread over time. In positions like the Nolot-class tests, humans are simply better. But OTB, where fatigue and distraction play a significant role, the computers have an edge.

Been that way for a long time, and nothing has really changed, except we are beginning to see the tactical frailness of human chess over a 3-4-5 hour game is the limiting factor that is making programs superior. But that is the _only_ reason...

A humans positional understanding is based on a plethera of 'rules of thumb' in which the majority have many exceptions. The computer is very good at finding the exceptions. The human needs to search inorder to find the exceptions. The computer is better at searching. There are holes in the positional undrstanding of even the strongest humans. The incredible searh of the modern chess program on modern hardware simply plugs alot of holes in its understanding. 10X? I don't buy it.

Positional judgement in GM-level players is not based on "rules of thumb" at all. It is based on a _large_ number of games played or analyzed, which teaches the GM which "patterns" are advantageous and which are not. A GM won't just stick a rook on the 7th rank like a program tends to do, unless their experience/understanding says it is good in this position. GM players play _very_ high quality chess at under 1 second per move, which means they look at only 1-2 moves in total (1-2 nodes). Try that with a program to see how badly they play if you limit them to even 1,000 nodes per search. Then you begin to see what the GM is doing that the computers are not yet capable of.

[bob]

Humans have _far_ superior positional understanding. By a factor of at least 10x and probably more than 100X better than any program around.

I agree about the 10X, but 100X? No way.
If you feed programs with a positional test suite you will notice that todays programs find a good percentage of the correct moves, much more than the past.

I would agree that perhaps it is for the wrong reasons, for example very deep tactics and not because they can plan for the far future as humans do when playing a positional move, but wrong or correct reasons, fact is they play many of the positional moves.

Also if you say that positional understanding of humans is about 10X to 100X better relative to computers, then i think you should emphasize the fact that computers are 500X to 1000X times better in all kind of tactics.

I believe I said exactly that by implication... Although given the choice of trusting a computer's 24 hour analysis or a strong GM's analysis, I would probably take the GM at 24 hours, they _still_ solve positions programs have no clue about..

Been that way for a long time, and nothing has really changed, except we are beginning to see the tactical frailness of human chess over a 3-4-5 hour game is the limiting factor that is making programs superior. But that is the _only_ reason...

I agree this is the only reason that top humans lose to computers these days, but i disagree that things has not changed for computers about playing positional moves.

Rybka for example, plays many amazing positional moves, mostly exchange type sacrifices and some times not because a 14 ply tactic gives back the material, but long term ones.
I remember a game where Rybka gave its Rook for a Bishop(not any Pawn) only to squeeze only to win 25 moves later the material. That was a true positional sacrifice. There are many such examples and not only with Rybka.

The problem is those are "general-rule sacrifices". And "general rules" have exceptions. I see many programs (mine included) make positional sacrifices that sometimes work out, but in other circumstances make me go "oh no, no way that works here because of ..."... and computers do that, including Rybka.

For example(i think this is a bad example):
Rybka here played the amazing 12...b3!! that wins.
[D]rnbqk2r/p3ppb1/3p3p/P1pP2pn/1pP1P3/5NB1/RP1N1PPP/3QKB1R b Kkq - 0 12

After the almost forced for white 13.Nxb3 Na6 14.Qb1 f5! 15.exf5 0-0 16.Bd3 Nb4 black is controlling the board.
Perhaps this 12...b3!! is a bad example of positional play and it's just a deep tactic, but in fact what we say positional play is just incredibly deep tactics!
Humans play some positional moves because they know(or they believe from experience) many moves later they will have the advantage.

While I agree that ultimately positional play is just deep tactics, the "positional play" is what humans still excel at, because the humans have a _far_ more powerful "computer" at their disposal. Not more powerful in terms of raw nodes per second, certainly, but the human brain is still not understood, much less emulated so far, and the level of parallelism and sophistication it has is incredible... And allows humans to make moves after searching just one move where a program would play beyond horribly searching that fast. Yes, I know the two players are "different" in many ways. And yes I know that _if_ chess is solved, then tactics rules and positional play no longer exists. But not today, not yet...

[Dr.Wael Deeb]

In the July 2008 edition of Chess Life (page 44), GM Lev Alburt
claims "Playing under classical conditions (40 moves in 2 1/2 hours),
today's best grandmasters are still favorites vs. Rybka and Fritz".

I was under the opposite impression, but I'm not a GM.

Sounds like a challenge for some program to me.

Just because someone is a GM, that does not mean he has a clue what he is talking about.
We want to see proofs, not baseless claims !
The top engines are proving their superiority more and more convincingly.

Matthias.

If you understood his point, the discussion would already be over.

Humans have _far_ superior positional understanding. By a factor of at least 10x and probably more than 100X better than any program around. however, the programs have far superior tactical skill, when the effort is spread over time. In positions like the Nolot-class tests, humans are simply better. But OTB, where fatigue and distraction play a significant role, the computers have an edge.

Been that way for a long time, and nothing has really changed, except we are beginning to see the tactical frailness of human chess over a 3-4-5 hour game is the limiting factor that is making programs superior. But that is the _only_ reason...

A humans positional understanding is based on a plethera of 'rules of thumb' in which the majority have many exceptions. The computer is very good at finding the exceptions. The human needs to search inorder to find the exceptions. The computer is better at searching. There are holes in the positional undrstanding of even the strongest humans. The incredible searh of the modern chess program on modern hardware simply plugs alot of holes in its understanding. 10X? I don't buy it.

Positional judgement in GM-level players is not based on "rules of thumb" at all. It is based on a _large_ number of games played or analyzed, which teaches the GM which "patterns" are advantageous and which are not. A GM won't just stick a rook on the 7th rank like a program tends to do, unless their experience/understanding says it is good in this position. GM players play _very_ high quality chess at under 1 second per move, which means they look at only 1-2 moves in total (1-2 nodes). Try that with a program to see how badly they play if you limit them to even 1,000 nodes per search. Then you begin to see what the GM is doing that the computers are not yet capable of.

You must be kidding us Robert....
Very high quality chess at 1 second per move,the humans
Even if they move the piece with an eye click,they won't be able to play a high quality chess,sorry,in this life time....