Neurosis 2.2

[Stan Arts]

Neurosis 2.2

Got tired of losing all the time, hate the holidayseason, so I decided to make my program more selective. Previously I tried to be as strong as possible with hardly being selective, but I got stuck more or less. Now I wrote some sort of a tactical ply that only tries most tactical moves after normal searchdepth, with a lot of threatdetections when not to and try a regular ply instead. If the tactical ply gives a big threat instead (gave check, moved a pawn to 7th, etc.) it can call a regular search ply again for the other side to try to refute the threat. That is more efficient then it sounds and basically my program now finds most tactics and some strategical ideas one ply faster, while often not losing a ply in searchdepth over a version without this, also it improves my effective branchingfactor so it works really well at long timecontrols. Maybe not very good at fast timecontrols, I dunno. Does cost me about 20-30% in nps. (~160Knps on PIV-2.4gHz now. )

Max hashtable size you can set increased from 64 to 128MB.

Mate-score-ranges changed, (First versions used 99.98/-99.98 for a mate in 1, but that's small if the user decides to insert a bunch of queens for example. Then I switched to 319.98/-319.98, and now to 299.98/-299.98) because of a new internal margin where I want to avoid checking not to cross the 16bit integer limit. Atleast Arena also announces the correct mate-in-N number with these scores.

Draw claims to Winboard by insufficient material added, (only K vs K and K+minor piece vs K) was never bothered about that but figured I'd add it sometime.

New endgame knowledge, tweaks and several speedups.

http://www.geocities.com/tasteofinsanit ... rosis.html
(Also on Leo Dijksman's WBEC website later I think. (Thx!) )

Stan

[Daniel Shawul]

Now I wrote some sort of a tactical ply that only tries most tactical moves after normal searchdepth, with a lot of threatdetections when not to and try a regular ply instead. If the tactical ply gives a big threat instead (gave check, moved a pawn to 7th, etc.) it can call a regular search ply again for the other side to try to refute the threat.

It sounds that you are skipping moves not reducing. If so is that done at nodes near to the root. My experience is that it can only be used neer the leaves.

Mate-score-ranges changed, (First versions used 99.98/-99.98 for a mate in 1, but that's small if the user decides to insert a bunch of queens for example. Then I switched to 319.98/-319.98, and now to 299.98/-299.98) because of a new internal margin where I want to avoid checking not to cross the 16bit integer limit. Atleast Arena also announces the correct mate-in-N number with these scores.

I also changed mate scores for some other reason. What I did was add a number different from one (mate + 100 instead of mate + 1). I am sure that no gui would display correct mate scores for this, so it might be better to let the engine send a mate string rather than trying it to handle itself.

regards
Daniel

[Stan Arts]

It sounds that you are skipping moves not reducing. If so is that done at nodes near to the root. My experience is that it can only be used neer the leaves.

Hi Daniel

Only at the leaves, otherwise it would indeed get too selective. For now it's always just one ply, and that means the risk is just one ply as well, that is one reason why I like it. Basically this is similair to reducing most quiet moves (and speeding up finding most tactical lines) at the leaves by 1 ply, but in reverse. (but that's just a matter of implementation)
I think it's an interesting way of searching and I hope to improve this further in the future. (by improving and adding threatdetections, and getting better at selecting moves to try)

I also changed mate scores for some other reason. What I did was add a number different from one (mate + 100 instead of mate + 1). I am sure that no gui would display correct mate scores for this, so it might be better to let the engine send a mate string rather than trying it to handle itself.

regards
Daniel

But now the gui-writers are on to you.

I have always sent a small <Mx> thing after the first move in my pv's indicating mate in how many, for example <M4> :
e4 <M4> e5 Bc4 Nc6 Qf3 b6 Qxf7#
But I also don't mind a gui to get the mate-in-N number right when it has that feature.

Scorpio is doing great in WBEC!

Regards
Stan

[Dann Corbit]

[[snip]
Scorpio is doing great in WBEC!

Code: Select all

Pl: Engine:           Score:
--- ----------------- ---------
 1: Rybka 2.2-x64     40.0 / 48
 2: Shredder 10-x64   33.5 / 48
 3: Spike 1.3x4       30.5 / 48
 4: Fruit 061027y     30.5 / 48
 5: Hiarcs X54-x64    29.0 / 48
 6: LoopMP 12.32      28.5 / 46
 7: Naum 2.1b-x64     27.0 / 44
 8: Ktulu 8.0         26.5 / 48
 9: Glaurung 1.21-x64 26.0 / 48
10: Scorpio 1.83      24.5 / 48
11: Gandalf 6.01      22.5 / 46 << PROFESSIONAL ENGINE
12: The King 3.33     22.5 / 48 << PROFESSIONAL ENGINE
13: DeepSjeng ZZ3     22.0 / 48 << PROFESSIONAL ENGINE
14: Ruffian 2.1.0     22.0 / 48 << PROFESSIONAL ENGINE 
15: SmarThink 1.0 uci 21.5 / 48 << PROFESSIONAL ENGINE 


I seem to remember a time when Daniel Shawul was totally disgusted with chess programming and ready to give it up. Now he's teaching the pros a few lessions.