Winboard Forum

by **Sune Fischer** » 27 Apr 2002, 13:25

Geschrieben von: / Posted by: Sune Fischer at 27 April 2002 14:25:37:
Hi
Some improvement over version 114 I hope

I ran a blitz match G/2 versus these engines:
Genesis v.1.03...........: 7,5 - 12,5 Frenzee v.117
TSCP v.1.73 (02/11/01)...: 3,0 - 17,0 -||-
Frenzee v.114............: 2,5 - 17,5 -||-
Gerbil...................: 11,5 - 8,5 -||-
It scores much better than the old version,
only lost by a narrow margin to Gerbil.
I don't have any idea how it will do in longer time controls though.
As usual, please report any problems to me.
I also played a match against Chezzz, but there it seemed it suddenly would lose on time in some endgames because it completely stopped drawing. I have no idea why that went wrong, but it was only against Chezzz. I am investigating, but haven't been able to reproduce it. I think maybe it is an Arena communication bug since I have played hundreds of games in winboard and never ever seen it there.
http://www.fys.ku.dk/~fischer/frenzee/frenzee.html
-S.

by **Uri Blass** » 27 Apr 2002, 22:36

Geschrieben von: / Posted by: Uri Blass at 27 April 2002 23:36:25:
Als Antwort auf: / As an answer to: New Frenzee version 117 geschrieben von: / posted by: Sune Fischer at 27 April 2002 14:25:37:

Hi
Some improvement over version 114 I hope
I ran a blitz match G/2 versus these engines:
Genesis v.1.03...........: 7,5 - 12,5 Frenzee v.117
TSCP v.1.73 (02/11/01)...: 3,0 - 17,0 -||-
Frenzee v.114............: 2,5 - 17,5 -||-
Gerbil...................: 11,5 - 8,5 -||-
It scores much better than the old version,
only lost by a narrow margin to Gerbil.
I don't have any idea how it will do in longer time controls though.
As usual, please report any problems to me.
I also played a match against Chezzz, but there it seemed it suddenly would lose on time in some endgames because it completely stopped drawing. I have no idea why that went wrong, but it was only against Chezzz. I am investigating, but haven't been able to reproduce it. I think maybe it is an Arena communication bug since I have played hundreds of games in winboard and never ever seen it there.
http://www.fys.ku.dk/~fischer/frenzee/frenzee.html
-S.

I have 2 questions:
1)How do I tell freenze the size of hash tables to use?
2)What source code did you read to learn about hash tables?
I may add hash tables to my program but I am very careful about big changes and I am not going to send a version with hash tables to a tournament without
testing it for at least one week including long time control games.
It means that movei may use hash tables in the 4th division only if I implement them in less than 2 weeks so I have at least one week for testing(I expect the 4th division to begin in less than 3 weeks).
Uri

by **Uri Blass** » 27 Apr 2002, 22:38

Geschrieben von: / Posted by: Uri Blass at 27 April 2002 23:38:52:
Als Antwort auf: / As an answer to: Re: New Frenzee version 117 geschrieben von: / posted by: Uri Blass at 27 April 2002 23:36:25:

1)How do I tell freenze the size of hash tables to use?

I mean frenzee(sorry about the spelling error)
Uri

by **Odd Gunnar Malin** » 27 Apr 2002, 23:16

Geschrieben von: / Posted by: Odd Gunnar Malin at 28 April 2002 00:16:51:
Als Antwort auf: / As an answer to: Re: New Frenzee version 117 geschrieben von: / posted by: Uri Blass at 27 April 2002 23:36:25:

I have 2 questions:
1)How do I tell freenze the size of hash tables to use?
2)What source code did you read to learn about hash tables?

Hi.
Try this:
http://www.breuker.demon.nl/thesis/index.html , Chapter 2
It discussing several methods.
Odd Gunnar

by **Sune Fischer** » 27 Apr 2002, 23:24

Geschrieben von: / Posted by: Sune Fischer at 28 April 2002 00:24:40:
Als Antwort auf: / As an answer to: Re: New Frenzee version 117 geschrieben von: / posted by: Uri Blass at 27 April 2002 23:36:25:

I have 2 questions:
1)How do I tell freenze the size of hash tables to use?
2)What source code did you read to learn about hash tables?
I may add hash tables to my program but I am very careful about big changes and I am not going to send a version with hash tables to a tournament without
testing it for at least one week including long time control games.
It means that movei may use hash tables in the 4th division only if I implement them in less than 2 weeks so I have at least one week for testing(I expect the 4th division to begin in less than 3 weeks).
Uri

its in the frenzee.ini file. also the readme.txt may be of some interest.
Bruce Moreland has a great tutorial on that hashing stuff:
http://www.seanet.com/~brucemo/topics/topics.htm
Be aware there is a problem with mate scores that he has left out, one should correct for the depth at which the mate is found, but you can check out how Crafty does that, it seems to work fine.
So far it is only using the upper and lower values, and will try the hashmove if there is one, but I'm having some problems with the exact score that is returned from q-search. when I enable it the searches slows down with many more nodes per ply, I don't understand that at all, so I'm playing without it for the time being.
Also not using killers yet, I haven't found an elegant way to implement them, so I'll way till I have studiet how Crafty and the others implement killers.
-S.

by **Andreas Herrmann** » 28 Apr 2002, 01:49

Geschrieben von: / Posted by: Andreas Herrmann at 28 April 2002 02:49:28:
Als Antwort auf: / As an answer to: Re: New Frenzee version 117 geschrieben von: / posted by: Sune Fischer at 28 April 2002 00:24:40:

I have 2 questions:
1)How do I tell freenze the size of hash tables to use?
2)What source code did you read to learn about hash tables?
I may add hash tables to my program but I am very careful about big changes and I am not going to send a version with hash tables to a tournament without
testing it for at least one week including long time control games.
It means that movei may use hash tables in the 4th division only if I implement them in less than 2 weeks so I have at least one week for testing(I expect the 4th division to begin in less than 3 weeks).
Uri
its in the frenzee.ini file. also the readme.txt may be of some interest.
Bruce Moreland has a great tutorial on that hashing stuff:
http://www.seanet.com/~brucemo/topics/topics.htm
Be aware there is a problem with mate scores that he has left out, one should correct for the depth at which the mate is found, but you can check out how Crafty does that, it seems to work fine.
So far it is only using the upper and lower values, and will try the hashmove if there is one, but I'm having some problems with the exact score that is returned from q-search. when I enable it the searches slows down with many more nodes per ply, I don't understand that at all, so I'm playing without it for the time being.
Also not using killers yet, I haven't found an elegant way to implement them, so I'll way till I have studiet how Crafty and the others implement killers.
-S.

Have you an extra hash table for the q-search or do you set a flag that indicates the q-search entrys? I think the best way is an extra hash table. So you don't overwrite q-search entrys with normal search entrys and reverse.
Andreas

by **Sune Fischer** » 28 Apr 2002, 06:38

Geschrieben von: / Posted by: Sune Fischer at 28 April 2002 07:38:28:
Als Antwort auf: / As an answer to: Re: New Frenzee version 117 geschrieben von: / posted by: Andreas Herrmann at 28 April 2002 02:49:28:

Have you an extra hash table for the q-search or do you set a flag that indicates the q-search entrys? I think the best way is an extra hash table. So you don't overwrite q-search entrys with normal search entrys and reverse.
Andreas

Okay, sounds interesting:)
Actually I do not even store the exact q-search values, because as I said it seems to spend more nodes at each ply that way.
I guess there is a bug somewhere, but currently I have no idea how to debug a hashtable.
-S.

by **Uri Blass** » 28 Apr 2002, 09:24

Geschrieben von: / Posted by: Uri Blass at 28 April 2002 10:24:16:
Als Antwort auf: / As an answer to: Re: New Frenzee version 117 geschrieben von: / posted by: Sune Fischer at 28 April 2002 00:24:40:

I have 2 questions:
1)How do I tell freenze the size of hash tables to use?
2)What source code did you read to learn about hash tables?
its in the frenzee.ini file. also the readme.txt may be of some interest.
Bruce Moreland has a great tutorial on that hashing stuff:
http://www.seanet.com/~brucemo/topics/topics.htm
Be aware there is a problem with mate scores that he has left out, one should correct for the depth at which the mate is found, but you can check out how Crafty does that, it seems to work fine.
So far it is only using the upper and lower values, and will try the hashmove if there is one, but I'm having some problems with the exact score that is returned from q-search. when I enable it the searches slows down with many more nodes per ply, I don't understand that at all, so I'm playing without it for the time being.
Also not using killers yet, I haven't found an elegant way to implement them, so I'll way till I have studiet how Crafty and the others implement killers.

I believe that it is a mistake to use hash tables in the qsearch.
Amir Ban explained in the chess computer club some years ago that Fritz5.32 and older version of Fritz did the same mistake and this is the reason that they needed big hash tables.
Things were changed after Amir talked with the programmer of Fritz Frans Morsch.

I studied how crafty implement killers and I thought that killers is more simple
to understand than hash tables.
I changed the structure of my move ordering function for it to do it similiar to crafty.
Uri

by **Andreas Herrmann** » 28 Apr 2002, 09:34

Geschrieben von: / Posted by: Andreas Herrmann at 28 April 2002 10:34:52:
Als Antwort auf: / As an answer to: Re: New Frenzee version 117 geschrieben von: / posted by: Uri Blass at 28 April 2002 10:24:16:

I have 2 questions:
1)How do I tell freenze the size of hash tables to use?
2)What source code did you read to learn about hash tables?
its in the frenzee.ini file. also the readme.txt may be of some interest.
Bruce Moreland has a great tutorial on that hashing stuff:
http://www.seanet.com/~brucemo/topics/topics.htm
Be aware there is a problem with mate scores that he has left out, one should correct for the depth at which the mate is found, but you can check out how Crafty does that, it seems to work fine.
So far it is only using the upper and lower values, and will try the hashmove if there is one, but I'm having some problems with the exact score that is returned from q-search. when I enable it the searches slows down with many more nodes per ply, I don't understand that at all, so I'm playing without it for the time being.
Also not using killers yet, I haven't found an elegant way to implement them, so I'll way till I have studiet how Crafty and the others implement killers.
I believe that it is a mistake to use hash tables in the qsearch.
Amir Ban explained in the chess computer club some years ago that Fritz5.32 and older version of Fritz did the same mistake and this is the reason that they needed big hash tables.
Things were changed after Amir talked with the programmer of Fritz Frans Morsch.

I studied how crafty implement killers and I thought that killers is more simple
to understand than hash tables.
I changed the structure of my move ordering function for it to do it similiar to crafty.
Uri

Hi,
Holmes writes at at the moment no q-search entrys to the hash, but i have heard that some programmers are using an extra hash table for the q-search.
have a nice day
Andreas

Homepage of Holmes

by **Uri Blass** » 28 Apr 2002, 09:42

Geschrieben von: / Posted by: Uri Blass at 28 April 2002 10:42:34:
Als Antwort auf: / As an answer to: Re: New Frenzee version 117 geschrieben von: / posted by: Sune Fischer at 28 April 2002 00:24:40:

I have 2 questions:
1)How do I tell freenze the size of hash tables to use?
its in the frenzee.ini file. also the readme.txt may be of some interest.

I forgot to ask about it.
I see nothing in the frenzee.ini file about hash
Here is a copy of the frenzee.ini file:

Configuration file for the Frenzee engine.
Use opening book (yes/no): yes
Computer will call flag (yes/no): yes

Uri

by **Leo Dijksman** » 28 Apr 2002, 09:57

Geschrieben von: / Posted by: Leo Dijksman at 28 April 2002 10:57:07:
Als Antwort auf: / As an answer to: Re: New Frenzee version 117 geschrieben von: / posted by: Uri Blass at 28 April 2002 10:42:34:

I have 2 questions:
1)How do I tell freenze the size of hash tables to use?
its in the frenzee.ini file. also the readme.txt may be of some interest.
I forgot to ask about it.
I see nothing in the frenzee.ini file about hash
Here is a copy of the frenzee.ini file:

Configuration file for the Frenzee engine.
Use opening book (yes/no): yes
Computer will call flag (yes/no): yes

Uri

Hi Uri,
I think you looked at the .ini file of Frenzee v114, the ini i have with Frenzee v117 looks like this:
Configuration file for the Frenzee engine.
------------------------------------------
Use opening book (yes/no): yes
Main hashtable size (MB): 50
Pawn hashtable size (MB): 4
Computer will call flag (yes/no): yes
Choose sign of PV (regular/inversed): regular
Resign when score reaches: -6.30
Best wishes,
Leo.

Engine-List.

by **Sune Fischer** » 28 Apr 2002, 10:04

Geschrieben von: / Posted by: Sune Fischer at 28 April 2002 11:04:57:
Als Antwort auf: / As an answer to: Re: New Frenzee version 117 geschrieben von: / posted by: Uri Blass at 28 April 2002 10:24:16:

So far it is only using the upper and lower values, and will try the hashmove if there is one, but I'm having some problems with the exact score that is returned from q-search. when I enable it the searches slows down with many more nodes per ply, I don't understand that at all, so I'm playing without it for the time being.
Also not using killers yet, I haven't found an elegant way to implement them, so I'll way till I have studiet how Crafty and the others implement killers.
I believe that it is a mistake to use hash tables in the qsearch.
Amir Ban explained in the chess computer club some years ago that Fritz5.32 and older version of Fritz did the same mistake and this is the reason that they needed big hash tables.
Things were changed after Amir talked with the programmer of Fritz Frans Morsch.

I studied how crafty implement killers and I thought that killers is more simple
to understand than hash tables.
I changed the structure of my move ordering function for it to do it similiar to crafty.
Uri

Yes, I didn't mean hash the qsearch, only the exact scores that are returned from the qsearch.
I can think of a few reasons to hash in qsearch and at the leaf, and a few not to. Main reason is that it's added complexity and if there is only a minimal effect or none at all, then I'd rather not do it.
If the eval() is heavy enough it may be worth it, one could do a seperate table just for leaf nodes with the simple replacement scheme of always to replace.
No flags to keep track of, just a score. Of course lazy eval and all that complicates things.
I just had a peek at that also, very simple actually, only real problem is how not to research the killers if they fail.
Looks like Crafty will reset the tried moves to zero, then crafty loops over the movelist and finds the best move that is not a zero-move.
It just seems a bit slow with all those checks if the node fails low.

by **Sune Fischer** » 28 Apr 2002, 10:18

Geschrieben von: / Posted by: Sune Fischer at 28 April 2002 11:18:05:
Als Antwort auf: / As an answer to: Re: New Frenzee version 117 geschrieben von: / posted by: Leo Dijksman at 28 April 2002 10:57:07:

Hi Uri,
I think you looked at the .ini file of Frenzee v114, the ini i have with Frenzee v117 looks like this:

just delete the old ini file, it will print a new one on the first run, with default parameters if it can't locate an ini file. It will always use those default values if it doesn't find something in the ini that tells it otherwise.
How do others do it, if this is not normal, I don't mind changing it.
I just haven't seen winboard pass any information about size of the hash.
-S.

by **Günther Simon** » 28 Apr 2002, 12:30

Geschrieben von: / Posted by: Günther Simon at 28 April 2002 13:30:41:
Als Antwort auf: / As an answer to: Re: New Frenzee version 117 geschrieben von: / posted by: Sune Fischer at 28 April 2002 11:18:05:

Hi Uri,
I think you looked at the .ini file of Frenzee v114, the ini i have with Frenzee v117 looks like this:
just delete the old ini file, it will print a new one on the first run, with default parameters if it can't locate an ini file. It will always use those default values if it doesn't find something in the ini that tells it otherwise.
How do others do it, if this is not normal, I don't mind changing it.
I just haven't seen winboard pass any information about size of the hash.
-S.

I think this is pretty normal and ok but how did you work out
this exactly resign value of 6.30?

)
Regards,
Günther

by **Sune Fischer** » 28 Apr 2002, 12:47

Geschrieben von: / Posted by: Sune Fischer at 28 April 2002 13:47:25:
Als Antwort auf: / As an answer to: Re: New Frenzee version 117 geschrieben von: / posted by: Günther Simon at 28 April 2002 13:30:41:

Hi Uri,
I think you looked at the .ini file of Frenzee v114, the ini i have with Frenzee v117 looks like this:
just delete the old ini file, it will print a new one on the first run, with default parameters if it can't locate an ini file. It will always use those default values if it doesn't find something in the ini that tells it otherwise.
How do others do it, if this is not normal, I don't mind changing it.
I just haven't seen winboard pass any information about size of the hash.
-S.

I think this is pretty normal and ok but how did you work out
this exactly resign value of 6.30? )
Regards,
Günther

It was a lot of hard science, you will have to wait for my thesis on that one

Winboard Forum

Who is online