>
> DF: how many ply is your usual search? I'm getting 3 to 6 ply in the main
> search, but the quiescence search often adds another 3 to 8 ply. The Q
> search has almost all the nodes.
>
> Right now I use weak patterns to "veto" moves from the tree. If a
> move is below a given rating I don't even try it if I am on the last 4
> ply of the search
>
>
I want to clarify this. FooBar is currently doing fixed 3 ply search
and not looking at the time clock on CGOS time-controls. BUT, this
is for 9x9 games. It occurred to me that you might be talking about
19x19 games which I haven't tested.
I don't currently use any form of quies because my evaluation function
consists of N play-outs, where N might be around 100 - 200 play-outs or
so. It might still be useful to have a layer that deal with
difficult to resolve situations but with play-outs you don't need
quiescence the same way you would with a static evaluation function that
is relatively naive about tactics such as a chess style evaluation
function.
Sometimes FooBar has 2 or more minutes left at the end of a totally
played out game doing fixed 3 ply searches. A lot of the time is taken
up by just 3 or 4 early moves and most of the 3 ply searches return very
quickly - so FooBar would be doing a fair number of 4 ply searches in
the middle game if I had time-control. I'll fix this soon, hopefully
it will give a nice improvement - I hope to get about 50 ELO or more for
this. To get even time management I may have to sacrifice 2 or 3 of
the 3 ply searches to get a more 4 ply searches.
I can do 5 ply searches in about 85 seconds per move on average - it
varied from about 40 seconds to 200. I did this estimate with 10
middlegame positions. The branching factor seems to be about 3.5 per
iteration. I'm going to test this to 7 ply later.
*DEPTH TIME BF
----- ----- -----
2 2.2 -
3 7.9 3.6
4 24.3 3.1
5 84.7 3.5
*
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
