[Computer-go] quiescence in UCT search

Dave Dyer ddyer at real-me.net
Tue Nov 8 22:04:17 PST 2011


At 08:53 PM 11/8/2011, Michael Williams wrote:
>I don't see much point in pruning the nodes at the root like this because UCT selection at the root is a very small part of program execution.  Making it quicker will not matter.  And you can't do the same thing at other nodes because the root is the only node in which you have an idea how much effort will be directed through it.

UCT by itself assures that exponentially less time is spent below the
unproductive nodes, but exponentially less is not the same as negligible.

In the typical case, where all the nodes but one are eventually
killed off, you know early when the search is over; that is, that
the result will not change within the time allotted.  
Those savings are cumulative.  

However many nodes are in contention up to the very end, you have 
additional time (not wasted on the dead children) to spend on them.





More information about the Computer-go mailing list