| | Strelka / Houdini search secrets exposed | |
| | Author | Message |
---|
Bettina Admin
Posts : 914 Reputation : 3742 Join date : 2009-11-27 Location : Bettina ( Sobel ? ) speakeasy 77>> Up All Night 77 >> Chelsea 77 >> little venice, maida vale 77 >> ???? Wendy D Green & Adnan Khashoggi know .... DO YOU ? £10,000.00 reward for information leading to me . but i am NOT here !
| Subject: Strelka / Houdini search secrets exposed Wed Apr 25, 2012 9:32 pm | |
| [You must be registered and logged in to see this link.] [You must be registered and logged in to see this link.]
| Alex Newman wrote:
I disaassembled its search and found only minor changes compared to the Robbolito 0.085
Larry Kaufman wrote:
Since you have disassembled Houdini search, can you comment on those differences you found that might account for the huge strength increase from Robbolito to Houdini 1.5? Critter author Richard Vida did so about a year ago, but the differences he cited were not enough to account for the large strength increase, in my opinion. Perhaps his list was incomplete, or else the gains came more from eval, I don't know.
Alex Newman wrote: I was very disappointed when I saw the changes. I expected more. Maybe eval changes are more substantial. Search changes: - Obfuscate the score shown by engine (i guess to hide its origins, lol). - Call pre-quiesent search for depths below 9 (instead of 7 in Robbo). - Different calculation of null move depth reduction. - Null move verification search. - Slight changes in the couple of pruning numbers (most of the pruning margins are the same as in Robbo, only multiplied by 2). - Adjustable futility pruning margins based on the total material on board (Robbo has fixed numbers).
Adjust mate hash scores by ply.Detection of known endgames in which case Houdini skips the search and sets the score to endgame eval.
Larry Kaufman posts
Could you elaborate on the Adjustable futility numbers? Where they larger for the endgame or for the opening? We haven't found that to be helpful so far in Komodo. Also, are you sure there were no changes to Singular Extension? I have reason to believe that there were.
Larry
Alex Newman posts
Maybe many of those small changes added up.
By the way, I don't have anything against Houdard and what he did. I have problem with him lying about it.
There are some small changes in singular extensions: - They are done inside the move loop (similar to Stockfish), but this is just for the performance reasons. - They are skipped for king and pawn endgames. - Search depth is always DEPTH/2 instead of the MINIMUM(12, DEPTH / 2)
Pawn futility prune margins start with 124 for start position (actual margin is 248 but everything is *2 in houdini). Then it increases gradually to 166 for no pieces. Depth adjustment for margins in Robbo is 32 * DEPTH, in Houdini is 8 * DEPTH (actually 16*DEPTH, but again everything is *2). | _________________ Bettina (Sobel ? ) speakeasy 77>> Up all Night 77 >>Chelsea 77 >> little venice 77 /Maida Vale >>> ???? Wendy D Green & Adnan Kashoggi know .... DO YOU ? £2,000.00 reward for information leading to me . but i am NOT here ! | |
| | | Bettina Admin
Posts : 914 Reputation : 3742 Join date : 2009-11-27 Location : Bettina ( Sobel ? ) speakeasy 77>> Up All Night 77 >> Chelsea 77 >> little venice, maida vale 77 >> ???? Wendy D Green & Adnan Khashoggi know .... DO YOU ? £10,000.00 reward for information leading to me . but i am NOT here !
| Subject: continued.... Thu Apr 26, 2012 2:21 pm | |
|
|
|
|
| Alex Newman | There are some small changes in singular extensions:
- They are done inside the move loop (similar to Stockfish), but this is just for the performance reasons.
- They are skipped for king and pawn endgames.
- Search depth is always DEPTH/2 instead of the MINIMUM(12, DEPTH / 2)
Pawn futility prune margins start with 124 for start position (actual margin is 248 but everything is *2 in houdini). Then it increases gradually to 166 for no pieces.
Depth adjustment for margins in Robbo is 32 * DEPTH, in Houdini is 8 * DEPTH (actually 16*DEPTH, but again everything is *2). | Larry Kaufman
Thanks. Regarding singular extension I'm asking about minimum depth requirement. In Robbo I believe it is 16 half ply for both PV and non-pv cases; in SF I believe it is 16 for nonPV but 12 for PV. What are the two numbers in Houdini? I suspect that at least one of them is 12 in Houdini 1.5.
Regarding futility margins my notes say Robbo uses 8*depth at low depth, and 48*depth at high depth, so I don't see where 32 is used? Can you clarify which margin you are talking about that uses 32? |
|
Alex Newman
|
|
|
| Depth condition in Houdini is always 16 just like Robbo.
Futility pruning code I am talking about is in Robbolito 0.0085g3 search_main.c at it looks like this:
if( VALUE >= POSITION_0->value + 75 + 32 * DEPTH )
{
NEXT->tag ^= bitboard_your_pawn; |
|
|
|
|
| Alex Newman | I was very disappointed when I saw the changes. I expected more. Maybe eval changes are more substantial.
Search changes:
- Obfuscate the score shown by engine (i guess to hide its origins, lol).
- Call pre-quiesent search for depths below 9 (instead of 7 in Robbo).
- Different calculation of null move depth reduction.
- Null move verification search.
- Slight changes in the couple of pruning numbers (most of the pruning margins are the same as in Robbo, only multiplied by 2).
- Adjustable futility pruning margins based on the total material on board (Robbo has fixed numbers).
- Adjust mate hash scores by ply.
- Detection of known endgames in which case Houdini skips the search and sets the score to endgame eval.
Maybe many of those small changes added up.
By the way, I don't have anything against Houdard and what he did. I have problem with him lying about it. |
Larry Kaufman Regarding this list, I believe it omits a couple of more significant changes that Richard Vida already talked about. One in particular was the idea of not going to quiescence at the usual depth<2 under certain circumstances. Can you verify this difference, and if so comment on the conditions for this "delayed quiescence"? |
|
Alex Newman
|
|
|
| I pulled the list from memory, so it's definitely not a complete list.
Yes, there is a delay to call quiescent in pv search under some conditions. If move played increases the danger to the king (as determined by the evaluation) and the move is a capture and eval score after move is greater then alpha+10 Houdini sets the search depth to 2 and doesn't call quiescent.
I don't know how this danger flag is set (it has only 3 values 0,1 and 2), because I didn't have a look at the evaluation ... yet. I am just guessing at this point that it represents some kind of danger level to the king.
There is also a counter that insures this extension is not done consecutively. |
|
_________________ Bettina (Sobel ? ) speakeasy 77>> Up all Night 77 >>Chelsea 77 >> little venice 77 /Maida Vale >>> ???? Wendy D Green & Adnan Kashoggi know .... DO YOU ? £2,000.00 reward for information leading to me . but i am NOT here ! | |
| | | Bettina Admin
Posts : 914 Reputation : 3742 Join date : 2009-11-27 Location : Bettina ( Sobel ? ) speakeasy 77>> Up All Night 77 >> Chelsea 77 >> little venice, maida vale 77 >> ???? Wendy D Green & Adnan Khashoggi know .... DO YOU ? £10,000.00 reward for information leading to me . but i am NOT here !
| Subject: Re: Strelka / Houdini search secrets exposed Thu Apr 26, 2012 3:21 pm | |
|
|
|
| Alex Newman
My guess is that all the small changes add up. Futility pruning change may have added 10 elo, null move change 15 elo, small pruning tweaks another 10-15 elo, contempt another 10 I tested tweaked material values from Houdini in the Robbolito and got around +10 elo.
Nothing remarkable. No imagination. No magic But good engineering job by Houdart. |
_________________ Bettina (Sobel ? ) speakeasy 77>> Up all Night 77 >>Chelsea 77 >> little venice 77 /Maida Vale >>> ???? Wendy D Green & Adnan Kashoggi know .... DO YOU ? £2,000.00 reward for information leading to me . but i am NOT here ! | |
| | | Bettina Admin
Posts : 914 Reputation : 3742 Join date : 2009-11-27 Location : Bettina ( Sobel ? ) speakeasy 77>> Up All Night 77 >> Chelsea 77 >> little venice, maida vale 77 >> ???? Wendy D Green & Adnan Khashoggi know .... DO YOU ? £10,000.00 reward for information leading to me . but i am NOT here !
| Subject: Re: Strelka / Houdini search secrets exposed Thu Apr 26, 2012 4:22 pm | |
|
|
|
|
|
Larry Kaufman
The futility change you mention only applies to pawn captures, and I'm pretty sure it's not worth more than 1 or 2 elo. "Small pruning tweaks" should not be worth anything like 10-15 elo, unless they are not really "small". Contempt I agree may be worth 10 elo against all the engines other than the elite ones. I'm particularly intrigued by your estimate of 15 for null move changes; can you describe those changes? They would have to be quite significant to be worth 15 elo.
Thanks!
Alex Newman
|
|
|
| Futility change is not done for pawns only. I used pawns just as an example.
Futility change applies to pre-quiescent and all quiescent searches, so the impact could be more significant.
I think generally null move verification is worth at least 10 elo. Robbo doesn't do any verification. Verification in Houdini is standard (depth-10 search). Nothing special.
I forgot to mention that Houdini also cancels reduced search (by returning beta-1) if null move returns a bad score (I think Stockfish does something similar).
There are 4-5 pruning tweaks (including delayed quiescent). I think 10-15 elo estimate is reasonable. |
|
|
|
-- nothing / no one is ever truly lost , just misplaced and awaiting us _________________ Bettina (Sobel ? ) speakeasy 77>> Up all Night 77 >>Chelsea 77 >> little venice 77 /Maida Vale >>> ???? Wendy D Green & Adnan Kashoggi know .... DO YOU ? £2,000.00 reward for information leading to me . but i am NOT here ! | |
| | | Bettina Admin
Posts : 914 Reputation : 3742 Join date : 2009-11-27 Location : Bettina ( Sobel ? ) speakeasy 77>> Up All Night 77 >> Chelsea 77 >> little venice, maida vale 77 >> ???? Wendy D Green & Adnan Khashoggi know .... DO YOU ? £10,000.00 reward for information leading to me . but i am NOT here !
| Subject: Re: Strelka / Houdini search secrets exposed Thu Apr 26, 2012 8:39 pm | |
|
| Larry Kaufman |
|
|
|
We have never found any measurable elo gain from null move verification, but maybe I'll take another look at that. The cancelled reduced search you mention could perhaps be significant. Can you mention any of the pruning tweaks that you deem significant enough to account for several elo? |
| Alex Newman
|
|
|
| I am too lazy to look for them now, but none of them are significant.
For instance in node_cut.c
VALUE + 35 + 2 * cnt
is changed to
VALUE + 44 + 2 * cnt |
|
Larry Kaufman
|
|
|
| Thanks. Regarding the cancelled reduced search idea, are you saying that any time the null move comes back below beta, LMR is cancelled? If so I just tried that in Komodo but it was clearly bad, though this could be due to redundancy with other things we do |
|
Alex Newman
|
|
|
| Make sure you clean reduced flag after lmr search returns.
You can look at Stockfish for some ideas about the criteria.
Houdini does it only when score is mate or less then beta-100. |
|
Alex Newman
|
|
|
| Verification search is done on higher depths (above 16 plies or so / half-moves ). So try using something like this:
if (depth-10 > 6)
{
turn off null move flag;
search at depth-10;
turn on null move flag;
if (score >= beta)
verification success;
else
verification failed;
}
|
|
_________________ Bettina (Sobel ? ) speakeasy 77>> Up all Night 77 >>Chelsea 77 >> little venice 77 /Maida Vale >>> ???? Wendy D Green & Adnan Kashoggi know .... DO YOU ? £2,000.00 reward for information leading to me . but i am NOT here ! | |
| | | Bettina Admin
Posts : 914 Reputation : 3742 Join date : 2009-11-27 Location : Bettina ( Sobel ? ) speakeasy 77>> Up All Night 77 >> Chelsea 77 >> little venice, maida vale 77 >> ???? Wendy D Green & Adnan Khashoggi know .... DO YOU ? £10,000.00 reward for information leading to me . but i am NOT here !
| Subject: Re: Strelka / Houdini search secrets exposed Fri Apr 27, 2012 4:06 pm | |
|
| [You must be registered and logged in to see this link.]
| Handbags at 10 paces .... pot calls kettle black [You must be registered and logged in to see this link.]
lkaufman wrote: You can't see the difference between relying on a verbal description of someone else's general ideas and copying an entire program verbatum? We are open to good ideas, whatever the source, but not to copying code.
Houdini wrote: Behind all your patronizing and moralizing facade, in this thread there's only one overriding thought in your interventions: "let's steal from Houdini".
Robert
Larry Kaufman wrote How can you complain about anyone "stealing" an idea or two from you, when you "stole" the entire code of Ippo/Ivanhoe?
--
| [You must be registered and logged in to see this link.]
| 2nds out round 2 [You must be registered and logged in to see this link.]
Houdart fires
let's look at the difference:
1) I did NOT "steal" from Ippolit or Ivanhoe because: - Ippolit is fully in the public domain, you cannot "steal" from it. - Houdini is 100% original at the source code level. - I have always clearly acknowledged the strong influence of Ippolit (it's still still on my web site, check it out).
2) You "steal" from Houdini because: - Houdini is NOT in the public domain. - I, as the author of the ideas that you would like to steal, strongly resent your act. - You incite other people to publish information obtained by RE'ing Houdini for the purpose of your commercial exploitation. In a lot of countries this is illegal.
Robert
******************* Bettina wrote - Houdini is 100% original at the source code level.
and Clinton swore he never had sex with Monica
--
| [You must be registered and logged in to see this link.]
| Alex Newman fire across houdarts bows [You must be registered and logged in to see this link.] Alex Newman wrote: I am pretty sure RE was and is an integral part of computer chess development from its beginnings. Being a target is just the price of being #1. just ask Vas :)
Houdart whinges
Obviously. But now we're in the new situation where the Talkchess forum is actively and publicly being used in the process.
What's next? Do I hire Yuri Osipov to completely RE Komodo 4 and to publish the Strelkomodo source codes on Talkchess?
| [You must be registered and logged in to see this link.]
| mouth allmighty get his little Napoleon dress on [You must be registered and logged in to see this link.] Julien MARCEL
[moderation] Openly discussing/publishing code obtained through reverse engineering is illegal. That part of the discussion has thus been erased. Please do not do that again in the future. [/moderation] __
--
| [You must be registered and logged in to see this link.]
| biggest cheater blows xx's and hot air. hppo hypocrisy [You must be registered and logged in to see this link.] Thank you Julie xxxx
Larry, I hope that you reflect on what was written - your attitude is not only morally wrong, but you are inciting people to illegal activities for your own commercial exploitation.
Robert
-- [You must be registered and logged in to see this link.]
Bettina
|
|
|
_________________ Bettina (Sobel ? ) speakeasy 77>> Up all Night 77 >>Chelsea 77 >> little venice 77 /Maida Vale >>> ???? Wendy D Green & Adnan Kashoggi know .... DO YOU ? £2,000.00 reward for information leading to me . but i am NOT here ! | |
| | | me4u Admin
Posts : 819 Reputation : 2737 Join date : 2010-01-05 Location : Pugatory
| Subject: Re: Strelka / Houdini search secrets exposed Fri Apr 27, 2012 4:14 pm | |
| - Bettina (Sobel ?) wrote:
| [You must be registered and logged in to see this link.]
| Handbags at 10 paces .... pot calls kettle black [You must be registered and logged in to see this link.]
lkaufman wrote: You can't see the difference between relying on a verbal description of someone else's general ideas and copying an entire program verbatum? We are open to good ideas, whatever the source, but not to copying code.
Houdini wrote: Behind all your patronizing and moralizing facade, in this thread there's only one overriding thought in your interventions: "let's steal from Houdini".
Robert
Larry Kaufman wrote How can you complain about anyone "stealing" an idea or two from you, when you "stole" the entire code of Ippo/Ivanhoe?
--
| [You must be registered and logged in to see this link.]
| 2nds out round 2 [You must be registered and logged in to see this link.]
Houdart fires
let's look at the difference:
1) I did NOT "steal" from Ippolit or Ivanhoe because: - Ippolit is fully in the public domain, you cannot "steal" from it. - Houdini is 100% original at the source code level. - I have always clearly acknowledged the strong influence of Ippolit (it's still still on my web site, check it out).
2) You "steal" from Houdini because: - Houdini is NOT in the public domain. - I, as the author of the ideas that you would like to steal, strongly resent your act. - You incite other people to publish information obtained by RE'ing Houdini for the purpose of your commercial exploitation. In a lot of countries this is illegal.
Robert
******************* Bettina wrote - Houdini is 100% original at the source code level.
and Clinton swore he never had sex with Monica
--
| [You must be registered and logged in to see this link.]
| Alex Newman fire across houdarts bows [You must be registered and logged in to see this link.] Alex Newman wrote: I am pretty sure RE was and is an integral part of computer chess development from its beginnings. Being a target is just the price of being #1. just ask Vas :)
Houdart whinges
Obviously. But now we're in the new situation where the Talkchess forum is actively and publicly being used in the process.
What's next? Do I hire Yuri Osipov to completely RE Komodo 4 and to publish the Strelkomodo source codes on Talkchess?
| [You must be registered and logged in to see this link.]
| mouth allmighty get his little Napoleon dress on [You must be registered and logged in to see this link.] Julien MARCEL
[moderation] Openly discussing/publishing code obtained through reverse engineering is illegal. That part of the discussion has thus been erased. Please do not do that again in the future. [/moderation] __
--
| [You must be registered and logged in to see this link.]
| biggest cheater blows xx's and hot air. hppo hypocrisy [You must be registered and logged in to see this link.] Thank you Julie xxxx
Larry, I hope that you reflect on what was written - your attitude is not only morally wrong, but you are inciting people to illegal activities for your own commercial exploitation.
Robert
-- [You must be registered and logged in to see this link.]
Bettina
|
|
|
Bettina wrote - Houdini is 100% original at the source code level. and Clinton swore he never had sex with Monica _________________ "No one is never ever truly gone as long as you remember them "
| |
| | | Doppelganger Dark Angels
Posts : 562 Reputation : 2159 Join date : 2010-02-03 Location : We often take for granted the very things that most deserve our gratitude. Cynthia Ozick
| Subject: Re: Strelka / Houdini search secrets exposed Sat Apr 28, 2012 8:51 am | |
| | |
| | | stevenaaus V.I.P.MEMBER
Posts : 12 Reputation : 62 Join date : 2012-01-14
| Subject: Re: Strelka / Houdini search secrets exposed Sun Apr 29, 2012 6:52 am | |
| Houdart is getting quite narky , laugh. - Quote :
- What's next?
Do I hire Yuri Osipov to completely RE Komodo 4 and to publish the Strelkomodo source codes on Talkchess? Hmmm... no | |
| | | Sponsored content
| Subject: Re: Strelka / Houdini search secrets exposed | |
| |
| | | | Strelka / Houdini search secrets exposed | |
|
Similar topics | |
|
| Permissions in this forum: | You cannot reply to topics in this forum
| |
| |
| |