Bullet masters. Pro Deo 1.0 - 2557.

[Kurt Utzinger]

Geschrieben von:/Posted by: Kurt Utzinger at 04 September 2004 21:08:08:
Als Antwort auf:/In reply to: Bullet masters. Pro Deo 1.0 - 2557. geschrieben von:/posted by: Igor Korshunov at 29. August 2004 13:17:

Bullet chess: can anybody really be interested
in this kind of chess
Kurt


Kurt & Rolf Chess

[Igor Korshunov]

Geschrieben von:/Posted by: Igor Korshunov at 04. September 2004 21:12:
Als Antwort auf:/In reply to: Re: Bullet masters. Pro Deo 1.0 - 2557. geschrieben von:/posted by: Albert Silver at 04 September 2004 19:48:44:

Because in more than 99% positions ANY engine will don't choose bad moves, if given enough time.
Exceptions only due to some heuristics like null move, because some engines can never find right move.
Follow your logic the most games must be draws, if reasonable time control used.
I CAN'T call such time control reasonable.
For me most reasonable time conrol where engines doing more mistakes.

That's highly questionable. Perhaps it depends on what you call bad moves.
But of course that's not the case. Perhaps you have some evidence that shows that there are fewer decisive games in 30 minute games than 2 minute games. I am testing some settings with Pro Deo against Junior 8 in 30 minute games using the Nunn 2 set. 13 games so far and only 2 draws.
Why? What does this prove?

Bad move - this is losing move.
It is very hard to find position where engine can't avoid lose, if it have enough time.
Yes. It is practically proven. Slower time control => more percent of draws.
I am talking not about several games, but about thousands.
Because time control which allowed engines make moves without mistakes totally useless.

I think this is a very primitive way of looking at chess with all due respect. I don't think computer chess is decided by tactical blunders by the respective opponents. I think it is much like grandmaster chess, in that a position is slowly built up and improved until a tactical sequence, if possible, appears on the board. For me, a bad move is a move that worsens the position, and not necessarily one that loses the game on the spot.
That is complete nonsense. I noticed that in the SSDF, Shredder 7.04 beat Ruffian 28-14. This was done at a time control of 40 moves in 2 hours, so obviously this wasn't enough time. How much time is necessary for Ruffian to stop losing so badly against Shredder?
If it's practically proven, it should be easy to point to evidence.
First, I find it amazing that anyone could think that engines don't make mistakes. Second, I can't even being to see the point of trying to find a time control where they make the most mistakes. You see, I never play bullet games against engines, so seeing the best bullet player is uninteresting to me, plus at bullet games, the engine will think for 1-2 seconds at most, if that, and I NEVER allow an engine to analyze a move for me for only 1-2 seconds. So if Ruffian can find the best moves in 1 second, and Pro Deo can find the best moves in 15-30 seconds or more, I will always analyze with Pro Deo since I couldn't care less about the results of 1 second of analysis.
Albert

In reality there are only wining, drawish and losing moves. There are no moves that can impove or worsen position. This moves are illusion.

I think that 100 years per move is bullet, because Ruffian will make a lot of mistakes with such time control.
I don't know how many time need Ruffian to play chess without many mistakes.

Of course, if you are not interested in bullet ratings, this time control is useless for you.

[Volker Pittlik]

Geschrieben von:/Posted by: Volker Pittlik at 04 September 2004 21:17:27:
Als Antwort auf:/In reply to: Re: Bullet masters. Pro Deo 1.0 - 2557. geschrieben von:/posted by: Igor Korshunov at 04. September 2004 21:12:

Hi Igor,
would you please request for a password? I think it is more comfortable for you and for us if we don't have to confirm 2 or 3 messages from you day by day.
Regards
Volker

[Albert Silver]

Geschrieben von:/Posted by: Albert Silver at 04 September 2004 21:21:26:
Als Antwort auf:/In reply to: Re: Bullet masters. Pro Deo 1.0 - 2557. geschrieben von:/posted by: Igor Korshunov at 04. September 2004 21:12:

It is very hard to find position where engine can't avoid lose, if it have enough time.

That is complete nonsense. I noticed that in the SSDF, Shredder 7.04 beat Ruffian 28-14. This was done at a time control of 40 moves in 2 hours, so obviously this wasn't enough time. How much time is necessary for Ruffian to stop losing so badly against Shredder?

I think that 100 years per move is bullet, because Ruffian will make a lot of mistakes with such time control.
I don't know how many time need Ruffian to play chess without many mistakes.

Maybe 100 years..
Anyhow, since at 40 moves in 2 hours it is still not able to avoid losing by a lot, it goes without saying that even 40 moves in 2 hours it is making many bad moves, so it is not too slow a time control.
Albert

[David Dahlem]

Geschrieben von:/Posted by: David Dahlem at 04 September 2004 21:24:07:
Als Antwort auf:/In reply to: Re: Bullet masters. Pro Deo 1.0 - 2557. geschrieben von:/posted by: Igor Korshunov at 04. September 2004 21:12:

Because in more than 99% positions ANY engine will don't choose bad moves, if given enough time.

That's highly questionable. Perhaps it depends on what you call bad moves.

Bad move - this is losing move.
It is very hard to find position where engine can't avoid lose, if it have enough time.

I think this is a very primitive way of looking at chess with all due respect. I don't think computer chess is decided by tactical blunders by the respective opponents. I think it is much like grandmaster chess, in that a position is slowly built up and improved until a tactical sequence, if possible, appears on the board. For me, a bad move is a move that worsens the position, and not necessarily one that loses the game on the spot.
That is complete nonsense. I noticed that in the SSDF, Shredder 7.04 beat Ruffian 28-14. This was done at a time control of 40 moves in 2 hours, so obviously this wasn't enough time. How much time is necessary for Ruffian to stop losing so badly against Shredder?

In reality there are only wining, drawish and losing moves. There are no moves that can impove or worsen position. This moves are illusion.
I think that 100 years per move is bullet, because Ruffian will make a lot of mistakes with such time control.

There is really no way to prove this, since your computer would be melted after thinking about a move for 100 years.
Dave

[Dan Honeycutt]

Geschrieben von:/Posted by: Dan Honeycutt at 04 September 2004 21:36:56:
Als Antwort auf:/In reply to: Re: Bullet masters. Pro Deo 1.0 - 2557. geschrieben von:/posted by: Albert Silver at 04 September 2004 21:03:41:

A similar study has already been done and published by Ernst Heinz on diminishing returns. This pretty much answers the question as it deals with the curve at which point greater depth (hence longer time controls) yields less and less gains. You can find it by searching the CCC archives.
Albert

Hi Albert:
Is this the same data that Uri posted a link to further down in the thread?
Best
Dan H

No, it is a different one. Here is the link to the thread from the CCC archives:
http://chessprogramming.org/cccsearch/c ... ead=112634
Albert

Thanks Albert. Interesting reading. If we start with:
(1) Long time control games tell you more than short time control games.
(2) More games tell you more than fewer games.
The ideal would be lots of games at long time control. But if you only have X time, is it better to play more games at shorter time or vice versa. Here begin heated debates, which I don't want to start. But the diminishing returns suggests that there is an optimum time control for giving the most information in a given amount of time. This, I believe, is what Volker Pittlik is working on with his short-long time control. Hopefully he will check in and give a progress report.
Dan H.

[Uri Blass]

Geschrieben von:/Posted by: Uri Blass at 04 September 2004 22:15:46:
Als Antwort auf:/In reply to: Re: Bullet masters. Pro Deo 1.0 - 2557. geschrieben von:/posted by: Kurt Utzinger at 04 September 2004 21:08:08:

Bullet chess: can anybody really be interested
in this kind of chess
Kurt

of course
I think that analyzing bullet games of programs in order to understand what are the mistakes that they do in these games can be a possible exercise for humans who want to improve their playing strength.
They can later compare their analysis with computer analysis.
I do not train in order to improve my playing strength but if I decide to train then it is a possible exercise to consider.
Uri

[Kurt Utzinger]

Geschrieben von:/Posted by: Kurt Utzinger at 04 September 2004 22:49:51:
Als Antwort auf:/In reply to: Re: Bullet masters. Pro Deo 1.0 - 2557. geschrieben von:/posted by: Uri Blass at 04 September 2004 22:15:46:

Bullet chess: can anybody really be interested
in this kind of chess
Kurt

of course
I think that analyzing bullet games of programs in order to understand what are the mistakes that they do in these games can be a possible exercise for humans who want to improve their playing strength.
They can later compare their analysis with computer analysis.
I do not train in order to improve my playing strength but if I decide to train then it is a possible exercise to consider.
Uri

I know a lot of people who play chess and with chess
computers but none of them has ever analysed bullet
games for learning purposes. And frankly spoken, I
simply can't imagine that such people exist.
Kurt

[Albert Silver]

Geschrieben von:/Posted by: Albert Silver at 05 September 2004 03:52:35:
Als Antwort auf:/In reply to: Re: Bullet masters. Pro Deo 1.0 - 2557. geschrieben von:/posted by: Kurt Utzinger at 04 September 2004 22:49:51:

Bullet chess: can anybody really be interested
in this kind of chess
Kurt

of course
I think that analyzing bullet games of programs in order to understand what are the mistakes that they do in these games can be a possible exercise for humans who want to improve their playing strength.
They can later compare their analysis with computer analysis.
I do not train in order to improve my playing strength but if I decide to train then it is a possible exercise to consider.
Uri

I know a lot of people who play chess and with chess
computers but none of them has ever analysed bullet
games for learning purposes. And frankly spoken, I
simply can't imagine that such people exist.
Kurt

I know one person, I'm afraid I must admit. It's true that here in Rio de Janeiro, he is a famous chess lunatic, behind many many hilarious stories. Anyhow, this guy literally WRITES down in a NOTEBOOK all of his blitz and 2-minute games to study and...(don't laugh)... admire. He will not hesitate to come to you, and say, I have a very interesting position I want to show you. I'd like your opinion. The guys is about 2150 or so. He'll then set it up, talk about its features, and if you're lucky, within 10 minutes he may mention it comes from some 2-minute bullet he played a week ago....
Albert

[Kurt Utzinger]

Geschrieben von:/Posted by: Kurt Utzinger at 05 September 2004 07:09:03:
Als Antwort auf:/In reply to: Re: Bullet masters. Pro Deo 1.0 - 2557. geschrieben von:/posted by: Albert Silver at 05 September 2004 03:52:35:

Bullet chess: can anybody really be interested
in this kind of chess
Kurt

of course
I think that analyzing bullet games of programs in order to understand what are the mistakes that they do in these games can be a possible exercise for humans who want to improve their playing strength.
They can later compare their analysis with computer analysis.
I do not train in order to improve my playing strength but if I decide to train then it is a possible exercise to consider.
Uri

I know a lot of people who play chess and with chess
computers but none of them has ever analysed bullet
games for learning purposes. And frankly spoken, I
simply can't imagine that such people exist.
Kurt

I know one person, I'm afraid I must admit. It's true that here in Rio de Janeiro, he is a famous chess lunatic, behind many many hilarious stories. Anyhow, this guy literally WRITES down in a NOTEBOOK all of his blitz and 2-minute games to study and...(don't laugh)... admire. He will not hesitate to come to you, and say, I have a very interesting position I want to show you. I'd like your opinion. The guys is about 2150 or so. He'll then set it up, talk about its features, and if you're lucky, within 10 minutes he may mention it comes from some 2-minute bullet he played a week ago....
Albert

Hi Albert
There is no rule without an exception
Kurt

[Albert Silver]

Geschrieben von:/Posted by: Albert Silver at 05 September 2004 20:18:46:
Als Antwort auf:/In reply to: Volker: How is the short-long time control doing? geschrieben von:/posted by: Dan Honeycutt at 04 September 2004 21:36:56:

A similar study has already been done and published by Ernst Heinz on diminishing returns. This pretty much answers the question as it deals with the curve at which point greater depth (hence longer time controls) yields less and less gains. You can find it by searching the CCC archives.
Albert

Hi Albert:
Is this the same data that Uri posted a link to further down in the thread?
Best
Dan H

No, it is a different one. Here is the link to the thread from the CCC archives:
http://chessprogramming.org/cccsearch/c ... ead=112634
Albert

Thanks Albert. Interesting reading. If we start with:
(1) Long time control games tell you more than short time control games.
(2) More games tell you more than fewer games.
The ideal would be lots of games at long time control. But if you only have X time, is it better to play more games at shorter time or vice versa. Here begin heated debates, which I don't want to start. But the diminishing returns suggests that there is an optimum time control for giving the most information in a given amount of time. This, I believe, is what Volker Pittlik is working on with his short-long time control. Hopefully he will check in and give a progress report.
Dan H.

I don't think the question is really that hard to answer if you consider the results of the diminishing returns and the well-known branching factors.
Today's hardware, even my medium-speed Athlon 2400XP+ (2 GHz), can get engines to a pretty good depth quite quickly. I'm testing Pro Deo settings and running Nunn matches to see the results. I've chosen 30 minute games as my basic testing parameter. Slow enough to be fairly deep, but fast enough to yield a reasonable amount of games. Perhaps not hundreds at a time, but I can already get a feel after 30-40 games to how it is playing and what it is doing wrong. I naturally compensate some of the lack of numbers by using my personal judgement.
Let's suppose for argument's sake that a top program has a branching factor of roughly 2.5. Let's also suppose that an ideal slow time control would be 40/2h or roughly 5 hours of play. On average. If I cut that in 2.5, reducing my average depth by a ply, the game would be in 2 hours. If I reduce this by another ply it would be in about 48 minutes. The 30 min KO games are about that if not a bit longer. The average depth I see in the middle game (no pondering) for Pro Deo is 12-13 plies. Sometimes it's more, and sometimes less. So this is very roughly 2 plies less than what a 40/2h game would reach. Since the returns seem to start flattening at around 12 plies or more (IIRC), this is not far from optimal. Always IMHO.
Albert