About LMR , again :-)

[Daniel Anulliero]

I want to make my LMR more sélective (and yes dangerous. .)
I know Senpai for exemple reduce by 2 at ply 6 and by ply/2 for remaining Depths
I want to try something :
I compute a certain % of the best history score and for exemple at Depth 6:
*reduction 4 ply if score < 10% of max history score
*reduction 3 ply if score < 20%
*else reduction by 2 ply
% increase with the plies (ply 18 for exemple : 20 and 53 %)
I noticed in my first tests it search more deeper but
I understand I must have a very good eval and move ordering of course...
to avoid too much blunders

[bob]

I want to make my LMR more sélective (and yes dangerous. .)
I know Senpai for exemple reduce by 2 at ply 6 and by ply/2 for remaining Depths
I want to try something :
I compute a certain % of the best history score and for exemple at Depth 6:
*reduction 4 ply if score < 10% of max history score
*reduction 3 ply if score < 20%
*else reduction by 2 ply
% increase with the plies (ply 18 for exemple : 20 and 53 %)
I noticed in my first tests it search more deeper but
I understand I must have a very good eval and move ordering of course...
to avoid too much blunders

What do you mean by "score < n% of max history score"???

[Daniel Anulliero]

Every moves have a score by the move history
(I have move history increase by (ply * ply))
I have a best score in the list
Then in lmr I do :
Every moves with score < 10% of the best score are reduced by 4ply
Every moves with score < 20% of the best score are reduced by 3ply
Else reduced by 2 ply
The percentages increase with the Depth
More clear I hope ..

Sorry it' s dificult to me to explain in english , pity I can't write in french here

[lauriet]

Have you done any testing ?
Before verses after ?

[bob]

Every moves have a score by the move history
(I have move history increase by (ply * ply))
I have a best score in the list
Then in lmr I do :
Every moves with score < 10% of the best score are reduced by 4ply
Every moves with score < 20% of the best score are reduced by 3ply
Else reduced by 2 ply
The percentages increase with the Depth
More clear I hope ..

Sorry it' s dificult to me to explain in english , pity I can't write in french here

If you wrote it in French, it wouldn't matter, as then I wouldn't understand it at all.

In any case, I am not a fan of using history counters like that. I have never had any luck with using them in any way other than the move with the highest history score is searched first...

[Daniel Anulliero]

Have you done any testing ?
Before verses after ?

Tests running. .. But not very easy , tuning the lmr need a good eval...I haven't yet...

[Daniel Anulliero]

Every moves have a score by the move history
(I have move history increase by (ply * ply))
I have a best score in the list
Then in lmr I do :
Every moves with score < 10% of the best score are reduced by 4ply
Every moves with score < 20% of the best score are reduced by 3ply
Else reduced by 2 ply
The percentages increase with the Depth
More clear I hope ..

Sorry it' s dificult to me to explain in english , pity I can't write in french here

If you wrote it in French, it wouldn't matter, as then I wouldn't understand it at all.

In any case, I am not a fan of using history counters like that. I have never had any luck with using them in any way other than the move with the highest history score is searched first...

Lol you know nothing about french ? Pff... and ALL the Web is in english ..argh
Ok thanks for now let's go to play with lmr !

[Daniel Anulliero]

well ...
Test was on the way but , many lost on time and others bugs stopped the test for a while
After many hours of debuging the test is really running now !
My LMR scheme:
I start to reduce when 4 moves full searched and depth start at 3
Do not reduce when :
*is in check
*give check
*PV move
*a pawn move at 7th rank
*move score > 70% of MAX_HISTORY

I'm testing 4 diférents versions , and my old engine Jars 1.75 for reference
each engine play 100 games vs each other at tc : 1'+1" (I know , may be too few games ) I can play more games in the future
Here is the crosstable at the moment :

Code: Select all

TEST LMR 1
DANY-PC, 2015.05.24 - 2015.05.25
Average Rating: 2130
                        Rtng    Score     
-------------------------------------------------------
 1: ISA lmr aggressif   2200  39.5 / 66   (+32 -18 =15)
 2: JARS_175            1850  35.0 / 65   (+31 -26 =8)
 3: ISA lmr depth - 2   2200  34.5 / 65   (+28 -24 =13)
 4: ISA pas de lmr      2200  29.0 / 66   (+24 -32 =10)
 5: ISA lmr a la senpai 2200  25.0 / 66   (+16 -31 =18)
-------------------------------------------------------
164 games: +62 -69 =32 

Code: Select all

versions tested :
ISA lmr aggressif &#58;	-reduction by 4plys if &#58; move searched >= 20 , depth >= 6 , move score < 15%¨of MAX_HISTORY
			-reduction by 3plys if &#58; move searched >= 10 , depth >= 4 , move score < 30% of MAX_HISTORY
			else &#58; reduction by 2 plys 
Jars 1.75 &#58; It's just my old engine for testing &#58;-)
ISA lmr depth -2  &#58;     -it reduce by 2plys just after the 4th moves  no reduced
ISA pas de lmr    &#58;  	-no LMR at all , for testing
ISA lmr a la senpai &#58;   -reduction by &#40;depth / 3&#41; plies but after 4 moves 

Of course , 160 games played is not suffiscent to give a good idea for the moment but there is some things interresting :
Strange is the "senpai" method behind the "no lmr" version
"lmr aggressive" is clearly ahead for the moment but the "classical " ply-2 is not so far
More disturbing is Jars not too far from ISA ... "lmr aggressive" score +9=2-6 against Jars for the moment
Not much improvement and some work again

To be continued ...

[Ferdy]

What about captures did you reduce it too?

[Daniel Anulliero]

Oops forgot that ...
Captures are not reduced of course ...