Quoting David Fotland <[EMAIL PROTECTED]>:
So I'm curious then. With simple UCT (no rave, no priors, no progressive
widening), many people said the best constant was about 0.45. What are the
new concepts that let you avoid the constant?
Is it RAVE, because the information gathered during the search lets you
focus the search accurately without the UCT term? Many people have said
that RAVE has no benefit for them.
Yes, it is RAVE, and mor specifil as it was last presented here
recently in the mailing list by the Mogo team, and not how it is was
originally presented in the mogo paper. Also there may be several
minor details that are peculiar to my implementation. Actually I did
not understand some aspects of the Mogo method mailed here and just
guessed some details. It suddenly worked and I could feel that the
search was unusually strong and selective, and since then I just
adjusted some parameters.
I used to do progressive widening but that is now turned off. RAVE is
free to pick any move that is not pruned right away.
Currently I believe that RAVE is only effective if one gets other
parameters right. For me it meant changing the uct parameter from 0.8
into 0.1. I also know of many pathological situations where Valkyria
currently will not find the best move, but rather the second best. It
is possible that other programs suffers even more than Valkyria from
similar problems and that this to some extent has to do with that the
nature of the playouts may interfere with AMAF. For example V either
plays forced moves or uniformly random among moves that are not
pruned. Other programs may rely on patterns to pick all moves in the
playouts and this might be bad for AMAF (this is a wild speculation).
Do most of the strongest programs use RAVE? I think from Crazystone's
papers, that it does not use RAVE. Gnugomc does not use rave.
You might not need it if you have strong pattern matching priors for
the tree part similar to Crazystone. RAVE makes it possible to ignore
most bad moves in a given positions. The weakness is that often some
good (with a chance of being the best possible move) are also ignored
completely.
Is it the prior values from go knowledge, like opening books, reading
tactics before the search etc? Do all of the top programs have opening
books now? I know mogo does.
Valkyria has just 4 moves in a hardcoded openingbook. Previous
versions used a book with several 1000's of positions that was both
self learned and modified by hand, but as long as the program changes
the book tend become inaccurate, so right now I do not use it and is
planning to write something more efficient than the old one which kept
each position as file on the harddrive.
Do most of the top programs read tactics before the search? I know Aya
does.
Valkyria only does some simple tactics in the playouts. It is stronger
than anything I ever programmed (on 9x9 at least) so currently I
cannot see how to integrate precomputed tactical results in the later
search. I think Aya is special because it was very strong doing search
before it went MC.
Does it matter how prior values are used to guide the search? I think mogo
uses prior knowledge to initialize the RAVE values. Do other programs
include it some other way, by initializing the FPU value, or by initializing
the UCT visits and confidence, or some extra, "prior" term in the equation?
Right know Valkyria sets priors for AMAF so that moves that are a good
local response to the last move have a prior 100% winrate with 20-100
visits depending on the priority of the triggered pattern. I think
Mogo has a fixed number of visits for the priorities but modifies the
winrate, but I never saw this described in a way that made it clear.
Previously I biased the UCT values after everyting else was computed
but found that this led to some bad behavior. By biasing the AMAF
values these biases will get less influential as the true winrate has
more weight than the AMAF-scores.
Are there other techniques (not RAVE) that people are using to get
information from the search to guide the move ordering? I think crazystone
estimates ownership of each point and uses it to set prior values in some
way.
I used to do that long time ago in Viking (the precursor to Valkyria)
that used alphabeta + MC-eval. As I remember it then it had a great
impact on move ordering that was quite bad (or even nonexistent) for
Viking.
I have tried it in Valkyria but was never able to see an improvement.
But I did not try hard enough to tell for sure. Both ownership and
AMAF use the same information (playouts), so trying to use it twice is
perhaps partially a waste of effort.
-Magnus
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
