Chesstempo VS chess.emrald

[richard]

Hi Boybawang,

I think that is an interesting idea and worth investigating.  I do foresee a few complexities. If the mate sequence is short , say mate in 2, then choosing the move from the previous ply may just make the computer look silly as the human can easily see that the mate is on the horizon and if the computer is not allowed to see that, may play what looks like a dumb move.  I imagine there will be situations where chosing the previous depth's line may not allow the tactic to play out properly, especially in situations where the opponent really only has one decent response once the opponent realising they are going to see a material loss on the current line. In fact it may lead to material actually being sacrificed earlier than was necessary, or in ways which make the tactic no longer required.  Anyway these concerns may not be an issue in practice and I do think the idea is worth consideration.

I'm planning a complete generator rewrite in the near future and when that is done I hope it will be easier to experiment with interesting ideas like this.

Regards,
Richard.

[boybawang]

Hi Richard,

It's not a problem!  The move from the previous ply is just the same as the current ply in most situation so the move won't look dumb.

Implementation might not be easy due to influence from Quiescence Search and Hash table.

You may be right on positions that doesn't include dropping of pieces.

A solution might be a material-balance-counter.  The idea is to simply discard variations with big material loss on its side. Just to get rid of dumb moves

Just an idea..

[richard]

Thanks boybawang,

I'll be investigating a few of these ideas once I get time to do the generator rewrite ( at the moment I still need to make updates to the current generator to make more short term problem set quality improvements, but I'll get to the rewrite eventually which will create a more flexible and more easily maintainable generator).

Regards,
Richard.

[boybawang]

with regards to "material imbalance accumulator"

The engines eval function can be slightly modified to prefer a defensive line that gains the most material despite being mated.

To the human solvers side, it's a variation with the best possible piece sacrifices in that position! Therefore, we will solve it brilliantly and satisfyingly rather than spoiled by ugly computerish defenses 

The line will look like a nicely composed one rather than dumb because it's still refined through Alpha Beta stage.

It just requires slight modification to the evaluation function:

1. All engine evals keep track of "total material imbalance".
In this case, we want that value to be as less as possible with a negative score because we want to sacrifice pieces.

2. All engine eval functions have MATE CONSTANT.  Mate constant is the biggest possible value the "evaluation function" can return which looks like 999.99 or so.

3. If it's a MATE, just subtract the "material imbalance" to the Mate Constant.
Therefore, the lesser the "material imbalance", the bigger the value returned by the "evaluation function".


Well, it's just my theory and I have doubts if it's flawed in practice 
I know you can do better than that richard...

[richard]

Hi boybawang,

Again some more interesting ideas :-)

I've modified toga's UCI interface in the past so I can ask for static material evaluations (for different purposes than the one we are talking about here).  In the end I decided to avoid direct engine hacking as it meant I couldn't easily switch between engines later if I wanted to, this means that I have to do a few things in my own code that the engine can do itself (for example looking at material balances in mating sequences where the UCI interface will only give me moves till mate, but for some purposes I need to know material counts so I implement my own).  I think some of what you are talking about could be implemented outside the engine, but as I said in a different post I need to rework the current generator to allow it to be more easily experimented with before I can start to play with some of these new ideas. I may also reconsider my current "don't touch the engine" approach if there are big gains in problem set quality to be gained by directly modifying the engine.

Regards,
Richard.

[texian]

i just stumbled across this old thread searching for something else and wanted to add my two cents .. my current CT/CTS ratings differences are in line with the rest of you (~1600 std/~1120) .. while the ratings formulas are similar, similarities end there .. in terms of difficulty, i'd say CT problems lead the way and my percentate solved confirs that 52%/69%

i started at CTS, then tried CT for only a few problems returning because i really like the black and white crisp board layout at CTS, but the clock is enough to drive me insane .. when i started doing CTS problems as if there were no clock,  i discovered i was learning more .. now that i've given CT another try, i do most of my tactics training here in super slow standard mode .. the variety of 3 and 4+ move problems here is much richer and larger .. now that i've slowed down even more, my ability to begin seeing the board and working out solutions is improving as well .. doing the hyper-blitz CTS problems, most 1 move problems at my level, that visualization aspect of my game was virtually non-existant and not really improving .. additionally, i can see an improvement in my games at FICS as can friends i play with regularly .. regardless, i like and reccommend both, but CT has become home and gets my 5 stars, CTS 4 stars .. once my rating improves a bit more, i look forward to getting the CT premium package

[allest]

I just found this site today, and I think it is much much better than CTS.  I think I actually emailed CTS or posted in their forums that I wish they had a standard or less time focused rating system used.  The problem with CTS is that if you are a standard player, and just want to do tactics to improve your standard play, you are greatly hindered by the rating system they use.  I could never get to the tougher problems, even though my accuracy was around 90+% just because I would take time to make sure I solved the problem accurately.  In standard, this is essential.  Whether you solve a tactical problem in 10 seconds or 30 seconds isnt really a big deal in standard time controls, whereas whether you solve the problem accurately or inaccurately is a big big deal . 
After 10 seconds or so, you may as well have solved the problem incorrectly rating wise, and you need to have a higher rating to get to harder tacitcal problems.  Hence, unless you want to solve a bunch of medium strength problems quickly and sloppily, you will never get to the harder problems on CTS.  This is terrible.
You could say that after solving medium problems enough, you should get to the point where you spot them instantly, but this isnt even helpful.  'If you see a good move, sit on your hands' -- Tarrash, Lasker, maybe some other people .  Even if you know your answer is right, for standard play, it is still better to take 10 seconds to look at other possibilities.  This way of tactical training is basically impossible on CTS, unless you want to solve the same 1400-1600 level problems over and over again.  It is unfortunate, because CTS could easily be as good as this site, but the super Blitz focus of the site really limits its general usefulness.

Thanks for making this site!

[drahacikfm]

Yes, absolutely.  If you make a move in 10 seconds in a new position you have never seen before, you are just guessing.  As far as training your brain to calculate variations, that is worse than useless.  It teaches you to be very sloppy, to guess, and to not calculate everything your opponent can do in reply.  Training on CTS will probably do very little to help your tournament chess.  It might help your online blitz play, but that's all.

That's why the untimed Standard mode here on Chess Tempo is so much better for improving your real chess (meaning not blitz).  You learn to really calculate lines and to be careful.

[uri blass]

Yes, absolutely.  If you make a move in 10 seconds in a new position you have never seen before, you are just guessing.  As far as training your brain to calculate variations, that is worse than useless.  It teaches you to be very sloppy, to guess, and to not calculate everything your opponent can do in reply.  Training on CTS will probably do very little to help your tournament chess.  It might help your online blitz play, but that's all.

That's why the untimed Standard mode here on Chess Tempo is so much better for improving your real chess (meaning not blitz).  You learn to really calculate lines and to be careful.

I discovered this thread only today when I decided to read more threads
so I respond only now.

I do not agree that if you make a move in 10 seconds in a new position you are just guessing.

It is clearly possible to see mate in 1 or mate in 2 without guessing in less than 10 seconds and there are other cases when you can be sure about the best move without guessing.

There are cases when you have to guess in CTS but if you know that this is a guess then you view the solution and think about it later so the fact that you did some guess in CTS does not mean that you are going to play this move in OTB game.

I can say that I have experience both in CT and in CTS.

In CTS I solved exactly 7500 problems and have today a new record of 1713
I also improved my blitz rating in this site and have again above 2200.

When I started in CTS I hardly got above 1500 because I did not guess enough and after the first 43 tries I had 40 problems correct out of 43 with rating of 1525

I still had 89% correctly even later after 214 problems when I got rating of 1601.
At that time I decided to guess more and improved  my rating to something near 1660 when my percentage dropped to 75.2%

At that time I decided that my target is to improve both my correct percentage and my rating and now I have 77.8% correct after 7500 tries with rating of 1723

Memorizing old problems is clearly one of the things that helps me to improve my rating and I decided not to guess a move when I have no idea if it works even if it can help my rating.

After some discussion in their forum when people said that it is harder to get big accuracy I decided to prove that I can do it with rating above 1400 so I opened a new handle and got 99 problems out of 100 correctly with CTS rating of 1416.

It proves that guessing did not cause me a demage in a game when the target is to have big accuracy(there were cases when I got the solution in less than 3 seconds and also cases when I needed more than a minute to check that there is no better move).

Uri

[chipschap]

once my rating improves a bit more, i look forward to getting the CT premium package

My rating is similar to yours and I already find the premium package really valuable in terms of tracking my results in many different ways.  The Gold package is more than worth the money.  Richard charges what seems like just a token amount and the value is very high.

[chipschap]

One other point in the CT / CTS comparison.  While it's only my opinion I believe the sense of community at CT is much stronger.  Although the total number of forum posters is small, I find them most supportive, helpful, and friendly.  I don't get anything like the same feeling at CTS.  CT is much like my home town chess club in its feeling.

[andreacoda]

While it's only my opinion I believe the sense of community at CT is much stronger. CT is much like my home town chess club in its feeling.

I totally second your point of view! I could not have said it better myself 
I LOVE this place!!

[dhingzpe]

What do you think about the relationship of my ratings on CT and CTS? My CT blitz rating is 21++, standard about 1800 while my CTS rating is about 1700.

[drahacikfm]

The rating systems are completely separate, so there isn't any reason the ratings should be close to each other.

The purpose of a rating system is to try to predict the result of one player against another (or a player against a problem).  The only relevant thing is the difference in rating between the two players.  For example, if you are rated 200 points higher than someone else, you are expected to score about 75% against that person (one win and one draw in two games).

If you added one million to the rating of every player and problem on CT or on CTS, the rating system would work exactly the same and just as well as it does now.  So you can't draw any conclusions from comparing ratings in separate isolated rating systems.

[revenant]

Chess News, Tuesday, January 19, 2038:  London:  "The World Chess Computer Championship round-robin tournament was won decisively today by Rybka 16.3 running at 0.8THz on a 256-CPU platform.  Coming in with a rating of 1,003,859, Rybka was favored to win over the rest of the field of 30 machines in which the closest competitor, Hydra 9.8, was rated a crippling 200 points lower at a measly 1,003,631.  Human grandmasters observing the tournament, the strongest of whom is currently rated about 3150, expressed continued bafflement at the tactics in some of the games, some of whose combinations ran 50 moves deep before the point became clear.  Programmers of the machines are pushing for a change in the rules of computer chess tournaments to make a 20x20 board size the standard so as to challenge the processors more."

"In other news, thousands of old desktop computers and embedded processors across the globe mysteriously crashed at exactly 3:14 a.m. this morning ..."

(little Unix in-joke there    )