Page 2 of 2

Re: alpha bound in hash table

Posted: Mon Nov 26, 2018 8:30 am
by xr_a_y
Tests I ran tonight show that storing alpha bound is very bad in terms of elo (-200).
I guess there is a mistake somewhere.
The actual mini source code

Code: Select all

 if ( skipMove==INVALIDMOVE && alphaUpdated) TT::setEntry({bestMove,bestScore,hashBound,depth,computeHash(p)});
 
removing the && alphaUpdated lose 200 elo ...

I am lost !

Re: alpha bound in hash table

Posted: Thu Nov 29, 2018 11:39 am
by RubiChess
xr_a_y wrote:
Mon Nov 26, 2018 8:30 am
Tests I ran tonight show that storing alpha bound is very bad in terms of elo (-200).
I will test how Rubi works without storing "bestmove not raising alpha" (which of course may not be the real best move) but I fear it won't win 200 ELO :-)

Andreas

Re: alpha bound in hash table

Posted: Fri Nov 30, 2018 6:53 am
by RubiChess
RubiChess wrote:
Thu Nov 29, 2018 11:39 am
xr_a_y wrote:
Mon Nov 26, 2018 8:30 am
Tests I ran tonight show that storing alpha bound is very bad in terms of elo (-200).
I will test how Rubi works without storing "bestmove not raising alpha" (which of course may not be the real best move) but I fear it won't win 200 ELO :-)

Andreas
Ran the test. Rubi loses about 10 elo if I don't store the "bestmove" at fail low.

Andreas

Re: alpha bound in hash table

Posted: Fri Nov 30, 2018 8:42 am
by xr_a_y
Ok thanks for the test. Re-running another myself with Minic...

Re: alpha bound in hash table

Posted: Sat Dec 01, 2018 8:33 am
by xr_a_y
xr_a_y wrote:
Fri Nov 30, 2018 8:42 am
Ok thanks for the test. Re-running another myself with Minic...
Unfortunatly I can confirm that in Minic, setting a alpha bound hash entry give very bad result, -200 elo in gauntlet with many other variations:

Code: Select all

   0 minic_dev                      95      15    1367   63.3%   37.2%
   ...
  12 minic_test                   -113      54      86   34.3%   47.7%
I'll run a 1versus1 selfplay to be sure...