I am also amazed. In my first MC program I played 100 PO for each of
those 50 candidate moves to even get a clue of where to search.
The playouts have a lot of go knowledge built in. If the last move
played has an effect tactically, Valkyria will almost always plays a
reasonable reply. When it plays a random move many moves are pruned
and a new random move is generated until an acceptable move is found.
If a white group has good eyeshape for example Valkyria will never
destroy that eyshape with bad random white moves.
The magic comes from AMAF I think, and it may be that AMAF in Valkyria
works even better with heavy playouts. Or maybe it is the case that
over time I have been forced to add tactical knowledge to Valkyria
exactly for those patterns that do not work well with AMAF.
Somehow it searches candidates implicitly with AMAF and tries to
refute the move that AMAF likes at the moment with real playouts. A
weakness is that in opening positions some moves are never searched
that could be potential good moves because of some bias caused by AMAF.
-Magnus
Quoting Fuming Wang <[email protected]>:
I am amazed at the rating that Valk3.5_100 got with just 100 playouts/move.
There are usually around 50 candidate move, and 100 playouts would give each
candidate just 2 playouts. It still amazes me how you can get reasonable
statistics with just 2 playouts. Statistically, this is unthinkable. I am
guessing that your playout engine has a lot of Go ability in itself, so that
every playout gives meaningful feedback, instead of relying on the
statistics of lots of playouts. Please enlight me.
Regards,
Fuming
On Wed, Sep 8, 2010 at 5:37 AM, <[email protected]> wrote:
Quoting Dave Dyer <[email protected]>:
But you have never (to my knowledge) layed out what way that is.
You're quite right here. I'm not advocating a specific change, just
pointing out that all the effort going into building faster monte carlo
engines may be irrelevant, because the programs actually need better
steering.
I have been working on Valkyria since 2006. Everytime I do something it
becomes slower. Meanwhile it has become about 1000 Elo points stronger (Only
200 Elo is due to faster computer). If you talk about people on running
their programs on as large clusters as possible, then I may agree, but
otherwise I think you misunderstand completely what people are doing to
improve their programs.
I know we disagree on this point, but I believe chess has reached it's
current state of success MOSTLY because of Moore's law.
It always was believed that Go was would have to be solved by other
means, perhaps even (gasp!) understanding the game. Monte carlo has
given some credibility to the theory that Moores law may be enough after
all. I'm arguing not.
Monte-carlo search *is* the "other means". Random exploration is exactly
what I do when I play go. The only difference is that my search is goal
directed so many playouts is just a 3-10 ply deep locally. As consequence I
am weaker than MC-program in actually evaluating the whole board position.
This weakness means I have to painfully compensate for it by counting
territory to set the ambition for the goals I search. I sometime have a
great intutions about playing some vital point. This caused by nothing else
but the human variant of AMAF.
Sure I do have a rich set of concepts that pop up in my thinking. But I am
afraid that this are just labels that I attach to my search results. I think
higher level concepts are very important for communicating about go, but
they are irrelevant for actually playing well.
The kind of knowledge about go that actally is essential for computers and
humans is the ability to play tactically correct quick and without error.
This means undrstanding L&D, seki, semeai, ladders and so on. And this is
also what makes Valkyria strong.
The reason Valkyria is not yet unbeatable is that the knowledge the
playouts have is still on a kyu level and very fragmented. There are
situations where I see the obious move in an instant where Valkyria needs to
search using several 100 playouts to get it right. In many cases it plays
perfect 100% of the time.
Get the fundmental knowledge right + MCTS = strong go
This has nothing to do with Moores Law. Valk3.5_100 is rated 1881 for 9x9
which is stronger than Gnugo. It only plays 100 playouts. When I started
doing MC evaluation with Viking5 in 2005 I had to spend 100000 playouts to
get close to beating gnugo.
Just another perspective
Magnus
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
