On Thu, Oct 01, 2009 at 04:50:26PM -0400, dhillism...@netscape.net wrote:
> For 9x9 games, when I added progressive widening to AntiGo (before I added
> RAVE), it was low hanging fruit. I used my old program Antbot9x9 for the move
> ranking and got a very nice strength increase for very little effort. Then,
> with a bit of tweaking, I got more improvement. RAVE, on the other hand,
> gives me a small improvement, and it comes at the cost of having to fuss with
> it endlessly.
I couldn't ever get RAVE to give any big improvement in the past. Two
things made big difference to me:
(i) Switching to the beta RAVE formula.
(ii) Making playouts much heavier - elaborate tactical check whether a
move is a self-atari, and full set of Mogo 3x3 patterns.
(iii) Collecting AMAF from the playouts too, but not from very late
playout stage (cutting out last 1/n of the playout and a simple
heuristic to just ignore nakade within playout proved to be equivalently
strong for me).
I think there is a lot of interaction between RAVE and the AMAF move
distribution your playouts give you. What is crucial is that (ii) may
give strong bias to some moves, but (a) it should still be slightly
randomized, not making the same move _always_ in the situation (b) it is
*CRUCIAL* that it does not _prohibit_ a move completely if it can be
good in any situation; it will not be explored in the tree search either.
My (otherwise already rather strong on 9x9) bot was not seeing snapbacks
in the actual game because I had a bug in my self-atari detection. My
self-atari detection is now extremely complicated to single-out the best
possible set of moves (including snapbacks, nakade moves, eye
falsification and false eye throw-ins); I would probably benefit from
simplifying it a lot, gaining much speed for the cost of sometimes
allowing nonsensical self-ataris.
If your policy never allows to make some move, try disabling that part.
--
Petr "Pasky" Baudis
A lot of people have my books on their bookshelves.
That's the problem, they need to read them. -- Don Knuth
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
