Better defense

[drahacikfm]

Another area where many problems could be improved is to give them more "human" defense.  By that, I mean that in many problems the computer just gives away material on the first move.  The computer could have held the material balance for 3 or 4 more moves, and then succumb to a mate or end up losing more material than it actually did with its "inhuman" first move.

The user's reaction is "why that stupid defense, I wanted to play the nice line I found!"  And the user doesn't have to show that he saw the point of the problem.

Seems like a very tricky thing to program.  But somehow favoring defenses that don't lose material immediately, even though they are not the computer's best move, might work.

[richard]

I agree that this is a bit of a problem but it is a problem that doesn't have any obvious solutions without completing changing the focus of the generator.  The favoring defenses that don't lose material would work for some positions, but as soon as you don't choose the best line you are risking choosing lines where the human will start saying, "The computer did WHAT!?".  At least at the moment the computers action's are justifiable. 

There is one way of dealing with this issue that would work and that is to change from a problem "generator" to a problem "verifier".  At the moment all the positions in the problem set start from a real position (the last move you see from the opponent is the move played in the actual game). However after that the tactics are what the computer sees, i.e. the tactic is generated from the computers suggested line.  Instead I could at that point not generate tactics  but look at the human moves and check that they match the computer analysis (and do all the other disambiguation  processing). I'm fairly sure this is what CTS does and it has a few advantages in terms of producing more human looking lines.  However the number of times you actually get a tactic to fully unfold in a real game , especially at the higher level is probably not that often, as the opponent will often see what is unfolding (at some point before all the material is gone) and resign.  CTS has more tactics that don't reach their natural conclusion for probably this reason. At the moment I'm hoping that after I get the generator to a level I'm happy with I can focus more on using comments and quality ratings to shape the set (I was hoping the last set would get there but the lingering alternative winning lines issues means it is not quite there yet).  If this shaping is done well then I can have the advantages of generation (access to larger pool of tactics and always having the full line available) and remove the disadvantages (computerish response lines and sometimes complicated lines "only a computer could see").

Regards,
Richard.

[pollock]

Have noticed in some problems the computer will immediately concede something rather than forcing me to demonstrate the tactical justification of that concession. Rather strange not having to demonstrate you spot the tactic. I cant see any good solution for this other than having a human go through problems. I don't mind it too much cause i can look at problem afterwards too see if i really understand.  - keep up good work 

[richard]

Hi Pollock,

The computer tends to assume the opponent also plays tactics like a computer :-)  Usually the concessions are simply the computer minimising long term material loss, it is not pretty chess but makes sense when you assume your opponent can also look at a million positions a second :-)  I agree that manual pruning is really the only way of dealing with these as we are really talking about aesthetics here which the computer isn't so good at..

Regards,
Richard.

[pollock]

Hi richard, I'm not sure its all aesthetics tho. Its practical chess that if your goner be a piece down or can enter a maze of complications where you risk being alot more down or escape you'd take the complications. Cause being down a piece vs a solid player is a guaranteed lose. Also ive seen this occur on some problems where i have played a simple move say a check or a capture which can only be played if understand a deeper tactical theme and then not have to demonstrated that theme cause the comp bales out early.

As mentioned before i only think this a minor issue and you clearly have many more important ideas and dreams for the site  . Maybe it would be more clear if i found some examples so ill keep an eye open for them  - pollock

[richard]

Thanks pollock,

I think I know the kind of problem you are talking about, but good examples are handy also.  I have a "hit list" of example problems that I try and remove/improve during each iteration of the problem generator.  This particular issues is likely not something I'll solve in the short term (and perhaps never in an automatic way), but having examples still helps..

Regards,
Richard.

[drahacikfm]

I've made comments on a lot of these problems, where the computer gives away material on move 1, when it's not clear why it has to.  It probably sees a mate that I don't see!  The comments contain "inhuman defense" or "non-human" defense, so if you can search comments for the word "human", you can find lots of these problems.

I still think it should be possible to automate this, and it's a very interesting artificial intelligence project.  Maybe if the second-best computer move keeps material equality for longer than the best move, then use that defense instead.
You could test on a set of problems that has my "inhuman defense" comments, and also include some normal problems, to see if the computer changes to some really stupid moves or not.

[drahacikfm]

I agree with Pollock.  Almost always, the "point" of the tactic is shown when the defender tries to maintain material equality for as long as possible.  And in a practical tournament game, if you have two choices, to give away a rook immediately and be completely lost, or to keep your rook and force your opponent to find a brilliant mate-in-5, you would choose to keep your rook.  The problems here should do that too, if at all possible.

[mrcm]

Richard, I think I found a way to make the computer solutions more "human".

The idea is that when the computer thinks for the defensive side, it must do it with a lot lower depth. This has the benefit of requiring less computer time, and if the computer is not allowed to think very deep, it sould not inmediatly give up material in order to avoid a very deep tactic that only could be found with a lot of plies.

That is...
 - Atacker thinks for 5 seconds computer time.
 - Defense thinks only 5 plies deep ( a few milisecons).  ( 5 plies or other thing similar)

[richard]

Drahacik: I'll have a bit more of a think about this one.  Almost definitely not for the next generator as I have some fairly obvious improvements I can make fairly quickly that I want to get started on and the generator set in motion.

Mrcm:  I certainly think this may create more human thinking in some positions, but I think it would be a double edged sword in many cases.  The danger of making the defender dumber than the attacker is that some tactics might be ruined by the opponent playing very weak defenses that make the nice tactic no longer required.  This would be especially true in longer move sequences but I guess I could dynamically choose the reduced horizon depending on tactic length. I think this idea might need some experimentation to better understand the pros and cons..I'll look at this in the generator improvements I'm planning after the next generator is done.

Another option might be to actually tweak the engine to explicitly prefer short term material equality more than it usually would. It is possible there might be a chess engine tuned for beating humans that does this already. I've tried to avoid engine modification in the past as if I rely on engine tweaks I lose the ability to switch between UCI engines which is handy.

Regards,
Richard.

[drahacikfm]

tama:  The computer didn't blunder.  Computers don't blunder on simple tactics.
You lose the knight even if you go to h3.

If you play Kh3, then PxB QxB Qh5+ and the Black queen is going to fork your King and knight no matter what you do.  For example, Kg3 Qf3+ Kh4 Qf2+ fork.  Or Kg2 Qe2+ fork.

[richard]

tama: As Drahacik points out it is very unlikely that the computer blunders (or at least blunders in a way that you will notice without being a GM).  The computer sometimes plays moves that a human would not play as it assumes the opponent will always find the best lines so sometimes it may give up material early on to avoid giving up more material later.  These can be hard to see as a human.

For your example of Kh3, as well as Drahacik's lines there is also h5 for black, white loses material trying to stop the mating threat on g5.

Regards,
Richard.

[tama]

Yes i know, after i analized Kh3 for 1 min i relized u still win the knight. Thats why i deleted my post Right away!

I was jumping to conclusions.

[tacto]

I know what you guys are going at and I understand and somewhat agree.

We just have to remember that the Toga 2 chess engine is probably stronger than any grandmaster in the world... ELO 2800 +- .

These inhuman defences are sometimes the "best" ones. It is true that they don't ever go with logic but only with calculation but we should look deeper into the problems before we scratch them.

Richard you mention something about using the problem Quality? I just want to make it very clear that a lot of users rate the problems low quality when they don't understand the position or when it's easy.

The quality should be based on ambiguity NOT brilliance and NOT difficulty. These people are ignorant and the who purpose of the quality function is probably - sorry to say - spoiled.

[richard]

Tacto, I agree on the problem quality ratings.  Right now, I could never just say, "remove all 1 star problems" as a number of them are quite good (and there are a few 5 star problems which are quite bad).  However they are quite useful in guiding where to first look for bad problems.   I also have some statistics which show me quality ratings based on the rating of the user making the assessment that I use to track overall problem set quality.

Regards,
Richard.