Winboard Forum

by **Heinz van Kempen** » 19 Jul 2004, 14:43

Geschrieben von:/Posted by: Heinz van Kempen at 19 July 2004 15:43:31:

Hi :-)

,
almost all votes are already given. Three more I am still waiting for. But some things are already decided. So even if all remaining ones are against learning the majority has already favoured that as well position learning and book learning should be allowed, (up to now only two against learning). But of course before starting we will reset all learning, because some top engines will have already learnt a lot from previous tournaments of the testers.
The other decisions for settings I will give tomorrow as well as the list of engines that got the most votes for both groups and will start the event, before others will be added. I think that I will have all the votes until then.
Very important: before starting all authors will have of course a few days to give us hints for settings or can also send us a new version. We decided to use also private engines (at least there was no protest against so far) so this version must not be a released one.
Best Regards
Heinz

by **Igor Gorelikov** » 19 Jul 2004, 15:34

Geschrieben von:/Posted by: Igor Gorelikov at 19 July 2004 16:34:37:
Als Antwort auf:/In reply to: learning---Amateur engines grand test geschrieben von:/posted by: Heinz van Kempen at 19 July 2004 15:43:31:

By my experience, learning has a small influence on performance (at least on recent level of its implementation). The main point of it is not repeating losing opening line over and over. I consider book and learning as an essensial part of an engine. So (IMHO) learning cannot be reset.
Igor

by **Bryan Hofmann** » 19 Jul 2004, 16:26

Geschrieben von:/Posted by: Bryan Hofmann at 19 July 2004 17:26:01:
Als Antwort auf:/In reply to: Re: learning---Amateur engines grand test geschrieben von:/posted by: Igor Gorelikov at 19 July 2004 16:34:37:

Actually the position learning is much more powerful then just learn of position not to repeat. Take for example the following position;
rnbqk2r/1p3ppp/p7/1NpPp3/QPP1P1n1/P4N2/4KbPP/R1B2B1R b - -
Crafty will make the move of Nd7
6 0.13 0.33 1. ... Nd7 2. Nd6+ Kf8 3. Ng5 Nb6 4.
Ndxf7 Qf6
6-> 0.24 0.33 1. ... Nd7 2. Nd6+ Kf8 3. Ng5 Nb6 4.
Ndxf7 Qf6
7 0.25 -1 1. ... Nd7
7 0.33 1.57 1. ... Nd7 2. Nd6+ Kf8 3. Nxc8 b5 4.
cxb5 Qxc8 5. bxa6 Rxa6
7-> 0.84 1.57 1. ... Nd7 2. Nd6+ Kf8 3. Nxc8 b5 4.
cxb5 Qxc8 5. bxa6 Rxa6
8 1.00 1.71 1. ... Nd7 2. Nd6+ Kf8 3. Nxc8 Ndf6
4. Qc2 Nh5 5. Nxe5 Nxe5 6. Kxf2 Rxc8
8-> 1.81 1.71 1. ... Nd7 2. Nd6+ Kf8 3. Nxc8 Ndf6
4. Qc2 Nh5 5. Nxe5 Nxe5 6. Kxf2 Rxc8
9 2.24 1.70 1. ... Nd7 2. Nd6+ Kf8 3. Nxc8 Ndf6
4. Qc2 cxb4 5. axb4 Nh5 6. Nxe5 Nxe5
7. Kxf2 Rxc8
9-> 5.45 1.70 1. ... Nd7 2. Nd6+ Kf8 3. Nxc8 Ndf6
4. Qc2 cxb4 5. axb4 Nh5 6. Nxe5 Nxe5
7. Kxf2 Rxc8
10 8.81 1.67 1. ... Nd7 2. d6 axb5 3. Qxa8 Nb6 4.
Qa5 Qxd6 5. h3 Bd4 6. Nxd4 Qxd4 7.
Qxb5+
10-> 13.11 1.67 1. ... Nd7 2. d6 axb5 3. Qxa8 Nb6 4.
Qa5 Qxd6 5. h3 Bd4 6. Nxd4 Qxd4 7.
Qxb5+
11 21.67 1.50 1. ... Nd7 2. Nd6+ Kf8 3. Ng5 Nb6 4.
Ndxf7 Nxa4 5. Nxd8 cxb4 6. axb4 Nb6
7. Ndf7 Rg8
But if I force it to learn that the following position is bad for white;
setb rnbqk2r/1p3ppp/8/1ppPp3/QPP1P1n1/P4N2/4KbPP/R1B2B1R w - -
store -32
Crafty will now make the move of axb5 instantly;
11-> 0.63 -32.00 1. ... axb5
12 0.63 -32.00 1. ... axb5
12-> 1.39 -32.00 1. ... axb5
13 1.39 -32.00 1. ... axb5
Because Crafty has learned that regardless of the score, it learned in the past that rnbqk2r/1p3ppp/8/1ppPp3/QPP1P1n1/P4N2/4KbPP/R1B2B1R w - - is a bad position for white so as black it will place white in this position. BTW for those that does not recognize this position it is the NoLot 6 position.

by **Igor Gorelikov** » 19 Jul 2004, 16:46

Geschrieben von:/Posted by: Igor Gorelikov at 19 July 2004 17:46:37:
Als Antwort auf:/In reply to: Re: learning---Amateur engines grand test geschrieben von:/posted by: Bryan Hofmann at 19 July 2004 17:26:01:

How many times this position can occur in engine to engine matches?
Igor

by **Uri Blass** » 19 Jul 2004, 16:54

Geschrieben von:/Posted by: Uri Blass at 19 July 2004 17:54:45:
Als Antwort auf:/In reply to: Re: learning---Amateur engines grand test geschrieben von:/posted by: Bryan Hofmann at 19 July 2004 17:26:01:

I know that crafty has antistmmetric evaluation so I wonder how it can learn from white about black unless you allow it also to learn wrong information?
Does it use learning only 1 ply after the root position or also later?
Does it learns also with some depth that if it can search deeper it does not trust the learning score?
Uri

by **Bryan Hofmann** » 19 Jul 2004, 17:01

Geschrieben von:/Posted by: Bryan Hofmann at 19 July 2004 18:01:53:
Als Antwort auf:/In reply to: Re: learning---Amateur engines grand test geschrieben von:/posted by: Igor Gorelikov at 19 July 2004 17:46:37:

This was just an example to show that the learning in crafty is not only for marking bad positions but if it can to place it's opponents in the bad position. If killer moves are made against Crafty it learns them and will repeat them in a game.

by **Bryan Hofmann** » 19 Jul 2004, 17:08

Geschrieben von:/Posted by: Bryan Hofmann at 19 July 2004 18:08:33:
Als Antwort auf:/In reply to: Re: learning---Amateur engines grand test geschrieben von:/posted by: Uri Blass at 19 July 2004 17:54:45:

It is not that it learns for wrong information, it learns that in a position it was bad or good. So say a engine made a killer move against crafty and crafty was white. So when it is playing black and has the chance to place white in the same position it will do such. That is why in my small test of Nunn I & II positions Crafty was able to do better as it had learned positions that Ruffian got Crafty into since the openings were mirrored.

by **Uri Blass** » 19 Jul 2004, 17:37

Geschrieben von:/Posted by: Uri Blass at 19 July 2004 18:37:12:
Als Antwort auf:/In reply to: Re: learning---Amateur engines grand test geschrieben von:/posted by: Bryan Hofmann at 19 July 2004 18:08:33:

But Crafty may evaluate position as bad for both sides because it has anti symmetric evaluation.
The fact that Crafty as black evaluate a position as +0.2 pawns for white does not mean that it should like to go for it as white because search may discover that the position is 0.2 for black if Crafty analyze it as white.

Uri

by **Bryan Hofmann** » 19 Jul 2004, 17:46

Geschrieben von:/Posted by: Bryan Hofmann at 19 July 2004 18:46:07:
Als Antwort auf:/In reply to: Re: learning---Amateur engines grand test geschrieben von:/posted by: Uri Blass at 19 July 2004 18:37:12:

You know where the code is.

by **Kurt Utzinger** » 19 Jul 2004, 18:24

Geschrieben von:/Posted by: Kurt Utzinger at 19 July 2004 19:24:09:
Als Antwort auf:/In reply to: Re: learning---Amateur engines grand test geschrieben von:/posted by: Igor Gorelikov at 19 July 2004 16:34:37:

In my opinion, learning is only justified in an
engine/engine-match but not in a tournament to
give all engines equal chances.
Kurt

Kurt & Rolf Chess

by **Andrew Fan** » 19 Jul 2004, 20:57

Geschrieben von:/Posted by: Andrew Fan at 19 July 2004 21:57:09:
Als Antwort auf:/In reply to: Re: learning---Amateur engines grand test geschrieben von:/posted by: Igor Gorelikov at 19 July 2004 16:34:37:

I also think learning should be allowed since we do it all the time by making new books off-line - with or without analysis. Engine learning is just an automated form of book editing (IMHO).

Andrew.

by **Uri Blass** » 19 Jul 2004, 21:43

Geschrieben von:/Posted by: Uri Blass at 19 July 2004 22:43:59:
Als Antwort auf:/In reply to: Re: learning---Amateur engines grand test geschrieben von:/posted by: Andrew Fan at 19 July 2004 21:57:09:

Who is "we"?
book editing is not done by everyone and I do not think that book update by the author during this tournament is allowed.
I do not think that we can generalize about engine learning and engines may also learn from games to change their evaluation or their search strategy.
I do not know about single engine that does it but the possibility always exists.
Uri

by **Andrew Fan** » 19 Jul 2004, 22:53

Geschrieben von:/Posted by: Andrew Fan at 19 July 2004 23:53:53:
Als Antwort auf:/In reply to: Re: learning---Amateur engines grand test geschrieben von:/posted by: Uri Blass at 19 July 2004 22:43:59:

"Opening Book":
A collection of positional information that is used to overcome the deficiencies of the human brain to transfer chess knowledge to machines - at least my brain seems deficient .
"We":
Chess players, engine authors, TD's, etc. Anyone who like to give good (or bad) guidance of opening lines that an engine should play.
"Learning":
Retriving previously computed, store or analysed data and use them constructively (or destructively). That included hash tables.

Any other suggestions?

Andrew.

by **Wael Deeb** » 19 Jul 2004, 23:05

Geschrieben von:/Posted by: Wael Deeb at 20 July 2004 00:05:39:
Als Antwort auf:/In reply to: Re: learning---Amateur engines grand test geschrieben von:/posted by: Kurt Utzinger at 19 July 2004 19:24:09:

In my opinion learning is an option which must be used no matter if the engine play eng.-eng. match or in tournament!If the other engines don't have this option,well....it's totally their problem!

by **Uri Blass** » 20 Jul 2004, 00:17

Geschrieben von:/Posted by: Uri Blass at 20 July 2004 01:17:59:
Als Antwort auf:/In reply to: Learning - my definitions geschrieben von:/posted by: Andrew Fan at 19 July 2004 23:53:53:

I did not ask about definitions
I think that we all agree that there is no problem with learning during the game(hash tables)
Majority also supported learning between games(book learning and positional learning) inspite of the fact that it means that the results are going to be less interesting for me.
The question was if to allow using learned information before the start of the tournament.
Heinz suggested to reset leasrning before the tournament.
You disagreed about it and I responded by doubting the reason that you gave that everyone does book editing.
I do not think that everyone does it and if people do it and edit the book of engines between games then it means that they test different thing than the thing that the author sent them.
Uri

by **Kurt Utzinger** » 20 Jul 2004, 06:56

Geschrieben von:/Posted by: Kurt Utzinger at 20 July 2004 07:56:53:
Als Antwort auf:/In reply to: Re: learning---Amateur engines grand test geschrieben von:/posted by: Wael Deeb at 20 July 2004 00:05:39:

I am going to think about this before starting
our long lasting AT 2004 (90m+30s) tournament.
Kurt

Kurt & Rolf Chess

by **Andrew Fan** » 20 Jul 2004, 15:47

Geschrieben von:/Posted by: Andrew Fan at 20 July 2004 16:47:54:
Als Antwort auf:/In reply to: Re: Learning - my definitions geschrieben von:/posted by: Uri Blass at 20 July 2004 01:17:59:

Ah! I finally see what the problem is - the words "we" and "all the time".
Believe it or not, English words are full of contradictions. "All the time" can mean "often", such as "I eat fish all the time" which really mean "I eat fish often".
I'll leave you to discover what the meaning of "we" is becuase this is off topic.

Andrew.

Winboard Forum

Who is online