I cannot comment on why Orego does things in a certain way, but you
adress things I guess make everybody doing UCT/MC scratch their heads.
Quoting Mark Boon <[EMAIL PROTECTED]>:
- It computes the UCT value in a completely different way. A comment
in the code refers to a paper called "Modification of UCT with
Patterns in Monte-Carlo Go". I haven't studied this yet, but whatever
it does it apparently doesn't do wonders over the standard C * sqrt (
(2*ln(N)) / (10*n) ) that I use.
Although MoGo may benefit from a modified UCT it does not mean that
light playouts do. With light playouts one has to search wide to
overcome strong bias in every position. With heavy playouts my theory
right now is that one can go much more seletive. And it also depends
on other tricks one is using. I am experimenting with Valkyria 3 right
know where C is optimal as 0.001 right know (but it is biased to
search wide initially by "assumed priors"-trick). In Valkyria 2 this
constant is 0.8 and earlier when the playouts were lighter I think it
was 1.8.
My guess is that what you are doing things right, and should be
careful adding stuff without testing. But later on you might benefit
from trying alternatives to UCT.
- It only initialises the list of untried moves in the tree after a
node had a minimum run-count of 81 (on 9x9). For the life of me I
couldn't figure out what the effect of this was or what it actually
does. I was wondering if this has an effect of what is counted as a
'run' but I'm not sure.
In Valkyria nodes are visited 7 times before expanding. Less and more
than that makes the program weaker. I would like it to be more because
it would save memory.
Then I found a paragraph (4.2) in Remi Coulomn's paper about ELO
raings in patterns. It briefly describes it as "As soon as a number of
simulations is equal to the number of points on the board, this node
is promoted to internal node, and pruning is applied." I can't help
feeling that the implementation in Orego is doing this. But I can't
figure out where it does any pruning or applying patterns of any kind.
Is there supposedly a general benefit to this even without pruning or
patterns? As stated before, at least it doesn't seem to provide any
benefit over my more primitive implementation. Maybe Peter Drake or
someone else familiar with Orego knows more about this?
I believe CrazyStone calculate "owner-at-end"statistics for each point
on the board before applaying patterns and determine the move order
for progressive widening. For the patterns it needs 64 runs to give
priority to points on the board that is unstable (and with a bias
towards defending own territory).
Anyway, reading the same paragraph mentioned above again I was struck
by another detail I thought surprising: after doing the required
number of runs, the candidates are pruned to a certain number 'n'
based on patterns. Does that mean from then on the win-ratio is
ignored? What if the by far most successful move so far does not match
any pattern? Am I misunderstanding something here? The paragraph is
very brief and does not elaborate much detail.
There are two possible ways of implementing the tree. One is with
lightweight nodes containg info only about the move and a pointer to
all children. The other one is with supernodes that contains the
information for all moves in a given position. In the former case the
information for first runs through the node is lost (actually the
children are random playout moves) in the latter case all information
is recorded and one should not prune winning nodes. What I mean here
is that there are two possible interpretations of what terminal node
might mean, and if Orego does one thing and you something else you are
bound to be confused.
In valkyria 3 I do move ordering as soon as a supernode is created.
But as I wrote above move ordering might become even better if it is
delayed about 50-100 runs if useful statistics can be collected from
the playouts other than the winrate itself.
On to my next step I introduced some very basic tactics to save stones
with one liberty, capture the opponent's stones with one liberty and
capturing the opponent's stones in a ladder. There are many possible
choices here. Just doing this near the last move and/or over the whole
board. Doing this in the simulation and/or during the selection.
I think you should only capture 1 liberty stones if the ladder is
broken. Likewise only save stones that can escape for sure. Perhaps
you did that but it sounded as if you only applied the ladder reading
code to capturing stones with two liberties.
My logic is that the program should only play deterministically when a
move is almost sure to be better than random tenuki. Capturing a
captured stone is in the worst case as bad as passing.
If you find out that I am wrong in the details here please tell me
because then I have to make a lot of changes to Valkyria... some of
these things I did out of mere conviction and I only know for sure
that the tactics used in total improves the playing strength.
Just doing this near the last move during simulation caused a slow-
down of a factor 4 or 5 but improves play considerably. Also doing
this near the last move during selection doesn't affect speed much but
deteriorated play! Doing this first near the last move and then look
for tactics over the whole board as a next step affected results
negatively even more. Number of playouts are still in the same ball-
park.
Thinking it over, since I don't use this to prune the selection but
just to order the candidates I could see that after many runs the
ordering suggested by the tactics get overriden by the UCT selection.
So I could see the effect of using this for selection reduced steadily
with the number of runs through a node. But still I didn't expect a
considerable reduction in strength. So what could be happening here?
I think I missed how you are using move ordering in you program. Basic
UCT will allocate one run to each child when a node is expanded. If it
follows the pseudcode on the Sensei Wiki-page then the move ordering
is random. If you just makes this part ordered then it will have
almost no effect on playing strength, because the move ordering will
only apply to one run for each child move.
If you are using progressive widening or bias the UCT-values with the
move order then you can have strong positive and/or negative effects.
If you force the program to allocate a lot of runs to capture dead
stones, then you lose a lot of runs for every dead block on the board.
In the worst case scenario on 19x19 in the endgame your program would
have tons of runs going to meaningless moves in the endgame.
I think my programs suffer from this to some extent I am aiming to
rethink how move ordering is done. In the last 19x19 tournament on
KGS, Remi pointed out that several of the weaker MC-programs seemed to
play a lot weaker as soon as the board was full with possible bad
captures.
- I could have a bug.
- I didn't run enough games (about 50)
- Using knowledge to order the initial selection is counter- productive
when not accompanied with pruning.
-Having a stronger program than Orego make me think you have no (less) bugs.
-50 games is very little unless you see a change of winrate over 15-20% or so.
But doing basic captures and saves should be a 30% increase in
strength against Gnugo for example. Actually I would claim it to be
the largest improvement you will ever see, because it is so
fundamental to the game.
-Pruning cannot be bad as long as bad moves are pruned. Too bad there
are so many moves that cannot easily be pruned...
The last one I find very hard to believe. Did anyone else run into
something like this?
Be glad that something works for you at least half of the time. I
reached the point where 9 out oft 10 changes to the program makes it
weaker...
Finally, I also looked a bit at using more threads to make use of more
than one processor. I figure this can wait and it's better to keep
things simple at this early stage but still it's something I want to
keep in mind. When looking at what I need to do to enable multiple
threads during search it seems to me I'll be required to lock
substantial parts of the UCT-tree. This means traversing the tree when
looking for the best node to expand is going to be the main
bottle-neck. Maybe not with just two to four processors, but I foresee
substantial diminishing returns after that. Is this correct? Is there
experience with many processors? Maybe a different expansion algorithm
will be required?
I think you are correct. What may make you happy is that every time
you make playouts slower you will also make the bottle-neck less of a
problem.
This also becomes an increasing problem with longer thinking times
because the UCT-tree can get really deep at least on small boards.
There is also a fundamental problem with diminsihing returns even if
the bottle-neck is neglible. With one thread the UCT algorithm picks
the "best" path through the tree at each run. When the results from
several threads are pending UCT will not be able to pick the "best"
path anymore and I think it is unavoidable that at some point the
benefit of adding more processors will not pay off the cost in $/Elo.
I am sure other programmers have better answers or different opinions
but I liked the topics you raised so I hope I was not to confusing in
my answer.
-Magnus
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/