>> [...] memory-saving technique of not expanding a leaf until you have
>> been through it many times (100, for example).
> 
> I have been wondering about this: If it pays off not to expand a node
> until it has been visited 100 times, why not bite the bullet and make
> those 100 simulations in one go? That should save a bit of time
> traversing the tree up and down. ...

My first reaction was there was good reason that waiting for 100
simulations. For instance if you get 10 losses out of the first 10
simulations for a node, then UCT will tend to ignore that variation, and
for good reason (probably). Doing another 90 simulations would be waste.

However, thinking about it some more, I wonder if the wait should be
based on the kind of results. E.g. if close to 0.5 (50% wins) then
expand (*) it after only 10 visits; if between 0.3 and 0.7 then expand
after 25 visits; otherwise wait for 100 visits. The idea being that when
random simulations are giving unreliable results you want that as a real
UCT node as soon as possible.

Darren

*: Either simply expand, or (for consistency) do the remaining
simulations in one go to get it up to 100, and then expand.

_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/

Reply via email to