Humans are better at positional chess, Corr. chess proves it

[Don]

What do you think would happen to a human master if playing against Houdini 3 on a multi core i7 computer.....Devastation!

The human master using H3 for analysis would most likely beat H3 quite easily at very long time controls. That is how correspondence chess works. And this is clearly what Leon Coleman meant.

That the human master + H3 beats H3 is clearly not because of the human master's tactical abilities.

The human probably doesn't even need to be close to master level.

Sorry that is not what he said.

"Pure engines playing correspondence chess against humans prove it."

He did not say pure engines playing against cyborgs prove it.

I play correspondence chess, that is not how correspondence chess works. Computer assistants must be stated to be allowed in the rules. No one wants to play against someone else's computer program, so a lot of event have a no computer rule.

But is your only argument is that humans must have Houdini 3 to beat Houdini 3. And you think that somehow this shows humans have better positional ideas on their own then Houdini 3. When they are stealing Houdini 3's own positional moves and analysis.

This is laughable....

No, they could have Stockfish and Chiron (just to name two), as the 3-Hirn experiments of Ingo Althoefer has shown.

Miguel

Lets play Miguel Your team of Stockfish and Chiron and you and you pick the move you want between the two programs Vs. Houdini 3.

All game analysis must be public, to be verified it was played by stockfish or Chiron and we use equal hardware.

Can we put money on this?

Lets play....

Mark, it's not a wise bet for you to make. However, I'd like to see it.

Ok I will bet a $1, but we agree on something. I want to see how two inferior programs beats a superior program because the human got to pick one of the inferior programs moves. I JUST GOT TO SEE THIS IN ACTION!

If this work like he claims I should make him play with sargon II and battle chess, Hell for the the fun of it he can use 3 programs, Crafty what ever version Miguel likes.

Mark,

In this case I don't agree with you. What you say seems to make logical sense on paper but it turns out that humans really do have some decision making skills that make them much better team members with a computer than sole players. Probably the number one skill we have is that we can integrate our opinions with the computers advice, a computer cannot do that very well. We are much better high level decision makers.

Imagine a player such a Carlsen (or any other top grandmaster) who had "Komodo built into their head" to get the picture. They would be making the decision but Komodo would be "advising" them. They would never make even a small positional or tactical error (as far as Komodo can judge) and they would also receive a second opinion and be alerted to possibilities they might have missed. Such a combination is more than the average of the skills and it's more than the strongest. It's not a chain that it only as strong as the weakest link either, it's more like the "whole that is greater than the sum of the parts."

A good metaphor for my example is to consider Carlsen the element of the team with superior management skills. The workers (the computer) may be superior to the manager, but they do not have management skills and a good manager is also an expert, but doesn't have to be the most competent - he just has to be a good manager. Humans are far better managers than computers.

The top players know more about the game by far than any chess program - they just don't play as well. In tennis the players will usually have coaches that have far more experience than the players being coached. They know more about the game than the players they coach and they are usually former great players with (hopefully) good coaching skills. But the coaches cannot beat the players they are coaching. What these coaches are doing is adding their experience and knowledge to the players they coach so that the whole is greater than the sum of the parts.

[velmarin]

New Komodo-Carlsen ???

[Dr.Wael Deeb]

New Komodo-Carlsen ???

[mwyoung]

I may not agree with me on this, but I am skeptical that their can be a 200 elo gain from this.

I have seen this system go very badly like in Carlsen vs the World, were the group of GM's picks the candidate moves, and the world voted on them. It did not work out so well for the wolrd team.

I am a tester, so I test this kind of stuff. I don't reject it out of hand. But I want to see data. So please help and vote.

I don't want to argue hypotheticals with people on CCC, I want DATA one way or the other.

http://talkchess.com/forum/viewtopic.ph ... 10&t=46156

[mwyoung]

Don, I have a question for you.

Miguel A. Ballicora, Ronald de Man and others theory is if I understand corrrectly is you can get a list of candidate moves say put Stockfish on pv=2,or 4. And from this list the humans pick the best move from a human point of view. This make the whole team stronger by 200 elo. In theory the stockfish team + human(s) are now stronger then Houdini 3.

And I guess you give this theory some merit...

My question is Miguel and Ronald seem to suggest this is proof humans are still better if the "theory is right" at positional chess, but you also seem to agree with this 3-Hirn teams theory, but if I understand your position. You think this is not proof, because you seem to suggest computers are stronger positional.

Why?

[Don]

I may not agree with me on this, but I am skeptical that their can be a 200 elo gain from this.

I have seen this system go very badly like in Carlsen vs the World, were the group of GM's picks the candidate moves, and the world voted on them. It did not work out so well for the wolrd team.

That is not the same thing and if I let weak players vote on my move I would play much weaker.

If you look at this as the Grandmaster dragging the computer level down you are missing the point. The point is that the Grandmaster has the ability to enhance the result the computer could obtain on it's own.

If you look at many of the forum posts you may notice that people will post about some stupid move the machine made and there will be a discussion and then eventually a conclusion will be reached. But nobody ever says, "I guess the computer knows best, it is just something too far beyond us so this silly looking move must really be best."

I can tell you from my own experience in computer chess tournaments (before Komodo) that even though my program would kill me in chess, there were times when it "threatened" to play a losing move and I could spot it immediately. It created tense moments when at the last second the score would drop and my program would "wake up" and play the right move. And I'm not a strong player.

In those days I could have enhanced the programs results if I were given the power to over-ride the computer or simply had the power of veto over its' moves or time control. Many other times I would have happily forced it to move instantly knowing that the move it wanted to play was clearly best.

I am a tester, so I test this kind of stuff. I don't reject it out of hand. But I want to see data. So please help and vote.

I don't want to argue hypotheticals with people on CCC, I want DATA one way or the other.

http://talkchess.com/forum/viewtopic.ph ... 10&t=46156

[mwyoung]

"That is not the same thing and if I let weak players vote on my move I would play much weaker."

Don it clearly states human contorllers of average club strength. I don't want to split hairs, but average club play is between 1400 and 1500 USCF.

That is much weaker then the computer program and the human 1500 uscf player is a valid controller from the info from the web site.

That is what they are saying. From the web site.

Strength of the man-machine-team
= strength of the programs
+ 200 Elo points improvement
+/- 50 Elo points deviation.

This estimate should hold also for 3-Hirn teams with other human controllers of average club strength instead of me


http://www.minet.uni-jena.de/fakultaet/ ... Talke.html

[Don]

Don, I have a question for you.

Miguel A. Ballicora, Ronald de Man and others theory is if I understand corrrectly is you can get a list of candidate moves say put Stockfish on pv=2,or 4. And from this list the humans pick the best move from a human point of view. This make the whole team stronger by 200 elo. In theory the stockfish team + human(s) are now stronger then Houdini 3.

There are many ways to do this and don't vouch for each one, but what I'm saying is that one man and one machine working together will produce much better chess than either by themselves. I'm assuming that the human has access to a flexible GUI that let's him experiment and there are no constraints, the human can use the program in any way he see's fit.

And I guess you give this theory some merit...

My question is Miguel and Ronald seem to suggest this is proof humans are still better if the "theory is right" at positional chess, but you also seem to agree with this 3-Hirn teams theory, but if I understand your position. You think this is not proof, because you seem to suggest computers are stronger positional.

Why?

There are good reasons why computers are stronger than GM's with positional play.

First of all modern evaluation functions are not as crappy as the critics want to make them out to be. Komodo has many terms and has encapsulated all the important principles of chess - and not just ad-hoc but in a balanced way. I admit it is not quite as flexible as a humans evaluation function when it comes to recognizing the exceptions but it's far better than the "critics" imagine it to be. It's probably even better in some area's than the top Grandmaster's because the relative difference in weights has been carefully researched and tested. Grandmasters do not even agree on what the relative importance of various chess concepts are and they all have biases - and so do programs.

But perhaps more to the point is that when you combine a high quality evaluation function with a search, you effectively multiply the power of it even to the point of simulating missing knowledge and putting incomplete knowledge back in balance. Not too long ago I explained how even a primitive evaluation function can simulate completely missing knowledge because of the pressure it puts on the search. For example if all you have is material evaluation and mobility, the program will still have some understanding of what a weak pawn is which will improve with depth because a weak pawn will force concessions the program DOES understand, such as actually losing the pawn or having to defend the pawn and thereby compromising the programs mobility. Of course it's always better to actually understand the principle directly, but this makes my point.

In actual practice you can use a computer to analyse the games of the top players and you realize that they make numerous minor errors. When you see such an error you have to determine if the human is in error or the computer or whether it's just a stylistic disagreement, but most of the time when there is a small but non-trivial score drop it's the human that made a small positional error.

I admit it's just a little tricky defining what positional play is. It's like a person who is physically ugly, it's hard to define it but we all know it when we see it. From the computers point of view there is nothing but tactics, maybe we should call it positional tactics. It's well known that humans make tactical errors much more than humans, but they also make positional error for similar reasons, not because of a lack of ability to judge the position but because they simply don't see that some positional concession is going to be imposed up on them - so these are in a way also miscalculations more than errors in positional judgement. But over the board it is indistinguishable from positional play. If you could transport these games back in time about 50-100 years ago before anyone knew about computers playing chess the top players would look at the games computers played and assume they were played by far superior positional players.

[Don]

"That is not the same thing and if I let weak players vote on my move I would play much weaker."

Don it clearly states human contorllers of average club strength. I don't want to split hairs, but average club play is between 1400 and 1500 USCF.

That is much weaker then the computer program and the human 1500 uscf player is a valid controller from the info from the web site.

That is what they are saying. From the web site.

Strength of the man-machine-team
= strength of the programs
+ 200 Elo points improvement
+/- 50 Elo points deviation.

This estimate should hold also for 3-Hirn teams with other human controllers of average club strength instead of me


http://www.minet.uni-jena.de/fakultaet/ ... Talke.html

That link is not working for me, but I agree that how the move is selected is an important issue that could change my conclusion. You probably missed my post which I just sent out - I am not sure of anything except that a human of moderate strength can use the computer as a tool to play better than either one by themselves. If you add more human players I am not sure that helps at all - I don't really know. I've played a lot of non-consultation chess at the club and it's great fun but it's clear that 2 players tend to work against each other. With consultation I feel that you still need the guy who makes the final decisions and simply takes other team members opinions into consideration. For example on a pure human consultation team the members can look for flaws in the plan or be used to explore or find new plans as long as each member is not fighting to play the game his own way. Voting can actually sets up an adversarial system.

If you do use voting for this type of decision making I would recommend the borda vote system. Each member would propose a set of candidate moves, they are combined, then the members rank each one from best to worst. The sum of the votes decides which move to play. This is better than a simple popular vote and the move selected will almost always be one that all members can live with, even if it's not any members favorite.

Don

[mwyoung]

Don, I have a question for you.

Miguel A. Ballicora, Ronald de Man and others theory is if I understand corrrectly is you can get a list of candidate moves say put Stockfish on pv=2,or 4. And from this list the humans pick the best move from a human point of view. This make the whole team stronger by 200 elo. In theory the stockfish team + human(s) are now stronger then Houdini 3.

There are many ways to do this and don't vouch for each one, but what I'm saying is that one man and one machine working together will produce much better chess than either by themselves. I'm assuming that the human has access to a flexible GUI that let's him experiment and there are no constraints, the human can use the program in any way he see's fit.

I agree, because I have done it. But why is this proof if a player uses rocking analysis to find better moves then a program could own its own show proof. That human have better positional knowledge then the computer alone has.

This has nothing to do with a persons better positional skills, they are just getting the most of the chess program by using the rocking analysis technique. The human operator will force the program to look deeper into a position by rocking the moves forward and back from a starting position in search of better deeper moves and filling up the hash tables. Then by trail and error extend out promising lines to reach a better conclusion about the position then the computers could by its own devices.

Yes, this works well for the human and computer team, but has nothing to do with human players superior positional skills.