Is search irrelevant when computing ahead of very big trees?
Posted: Wed Jul 24, 2013 11:19 am
Hello,
I have a though in my mind for some time I want to share to ask for opinions. It is about the eternal debate of reaching more search depth vs knowledge. The open question is: what is more important today: depth or knowledge? I will center the debate in long (very long) time controls.
My though is as follow: in a "lot of time calculated tree" from a equal position where two strong engines plays without obvious piece hanging in the short and middle term, there are always nodes in all paths where position is not tactical and it is 'calm', so we could draw another tree, with less depth, where all sheets are important positional points. My theory is that those positions is like a kind of style choosing set, and looking ahead is not futile, but maybe a waste of time. At the end, the engine will think there will be only one of those paths as the 'win' path, but the true is that more than one path are valid. In those cases, is more imporant an accurate eval, than a good search, as the time is supplying/hiding the search problems.
In the past, as there were not enought cpu power to complete a "tactic free" tree, this did not apply.
I can resume my though in two points:
1) In a match of equal strength engines, where one plays, for example, with 15 min/move, and the other 30 min/move, both will ends in similar strong result. I have not tested this, but suspect it could be true, as is the quality of the eval function what determine which engine is stronger.
2) I have personal subjetive evidence that invest in knowledge pays off more that in the search in the long run. With Rodin, I get more result for eval function than trying search improvements.
So the open question is: with enough time, is there a point in the search where computing further than that could be irrelevant?
Regards
Fermin
P.S. Maybe a complicated explanation from me, but hope you get the point.
I have a though in my mind for some time I want to share to ask for opinions. It is about the eternal debate of reaching more search depth vs knowledge. The open question is: what is more important today: depth or knowledge? I will center the debate in long (very long) time controls.
My though is as follow: in a "lot of time calculated tree" from a equal position where two strong engines plays without obvious piece hanging in the short and middle term, there are always nodes in all paths where position is not tactical and it is 'calm', so we could draw another tree, with less depth, where all sheets are important positional points. My theory is that those positions is like a kind of style choosing set, and looking ahead is not futile, but maybe a waste of time. At the end, the engine will think there will be only one of those paths as the 'win' path, but the true is that more than one path are valid. In those cases, is more imporant an accurate eval, than a good search, as the time is supplying/hiding the search problems.
In the past, as there were not enought cpu power to complete a "tactic free" tree, this did not apply.
I can resume my though in two points:
1) In a match of equal strength engines, where one plays, for example, with 15 min/move, and the other 30 min/move, both will ends in similar strong result. I have not tested this, but suspect it could be true, as is the quality of the eval function what determine which engine is stronger.
2) I have personal subjetive evidence that invest in knowledge pays off more that in the search in the long run. With Rodin, I get more result for eval function than trying search improvements.
So the open question is: with enough time, is there a point in the search where computing further than that could be irrelevant?
Regards
Fermin
P.S. Maybe a complicated explanation from me, but hope you get the point.