Another thing to take into consideration is the position specificity, I mean on equal positions with high probability of changing the best move several times you might get worse move ordering.
On sharp/winning positions you get much better move ordering.
But I suspect you already know that
[Discussion] - Measuring move ordering
Moderators: hgm, Rebel, chrisw
-
- Posts: 362
- Joined: Thu Mar 16, 2006 7:39 pm
- Location: Portugal
- Full name: Alvaro Cardoso
-
- Posts: 892
- Joined: Sun Nov 19, 2006 9:16 pm
- Location: Russia
Re: [Discussion] - Measuring move ordering
There is a true metric: ratio of nodes pruned relative to full minimax tree.
-
- Posts: 4366
- Joined: Fri Mar 10, 2006 5:23 am
- Location: http://www.arasanchess.org
Re: [Discussion] - Measuring move ordering
The stats I have are over the whole search tree (excluding qsearch), not just for the root. The character of the position will likely change as the search progresses deeper.
--Jon
--Jon
-
- Posts: 362
- Joined: Thu Mar 16, 2006 7:39 pm
- Location: Portugal
- Full name: Alvaro Cardoso
Re: [Discussion] - Measuring move ordering
Jon, my stats also come from the whole search tree except qsearch.
What I have observed is that an equal and unstable position (several changes in the best move) at the root usually gives poorer FH ratios like for example 82%8%2%1%.
And a winning stable position usually gives much better FH ratios, usually something like 95%2%1%0% and sometimes even something like 98%1%0%0%.
Notice I've used the word usually, there are many exceptions.
Engines are different of course, and mine is even more different since it is a checkers engine.
What I have observed is that an equal and unstable position (several changes in the best move) at the root usually gives poorer FH ratios like for example 82%8%2%1%.
And a winning stable position usually gives much better FH ratios, usually something like 95%2%1%0% and sometimes even something like 98%1%0%0%.
Notice I've used the word usually, there are many exceptions.
Engines are different of course, and mine is even more different since it is a checkers engine.
-
- Posts: 362
- Joined: Thu Mar 16, 2006 7:39 pm
- Location: Portugal
- Full name: Alvaro Cardoso
Re: [Discussion] - Measuring move ordering
I understand.
But these days with such agressive alphabeta LMR/LMP/Razoring/Futility, etc, the tree get so different from the minimax tree that is almost impossible to do that true metric you suggested.
But these days with such agressive alphabeta LMR/LMP/Razoring/Futility, etc, the tree get so different from the minimax tree that is almost impossible to do that true metric you suggested.