On Wed, Jun 29, 2011 at 10:26 PM, Peter Drake dr...@lclark.edu wrote:
On Jun 29, 2011, at 12:14 PM, Erik van der Werf wrote:
I hope you are aware that some strong MCTS programs use (at least) a
factor hundred less playouts to break even with gnugo. In fact, to get
to 50% they don't even need
On Jun 29, 2011, at 5:09 PM, Imran Hendley wrote:
Thanks for the detailed explanation of the paper.
Would it make sense to vary the number of moves generated by the
classifier as you run more playouts? Have you tried this? It seems
like the classifier would return garbage initially and
True. It performed better at the times we tried, and vanilla MCTS did
not appear close to catching up. In the theoretical limit, though,
MCTS is clearly a richer representation.
On Jun 30, 2011, at 7:35 AM, Erik van der Werf wrote:
On Wed, Jun 29, 2011 at 10:26 PM, Peter Drake
On 30-06-11 00:42, Lucas, Simon M wrote:
[...] For example, it sounds like those sizes are based
on all the entries that could possibly occur: perhaps in practice only a
relatively small
number of entries actually occur, and the statistics of those occurrences
can be
estimated
The general Monte Carlo approach is:
Repeat until golden brown:
Perform a playout, guided by the current policy
Determine the winner
Adjust the policy
The policy is adjusted so that winning moves are played more often,
losing moves less often (with some exploration
I attempted that many years ago. It was ~2004, IIRC, before UCT/RAVE.
I started from the simulated annealing model from Bruegmann's paper. Using
data from that search, I tried to create decision trees along the lines you
said.
The problem was that performance was terrible. The cost of building
I can't find the word local in the paper. Can you find the statement
you're referring to?
My mistake. In 4.1 it says, Moves were only considered if they were on the
3rd or 4th line or were within a large knightâs move of an existing stone.
I misread existing as previous somehow.
We tried looking at local patterns and at board locations in 3x3 or
large-knight's-move neighborhoods. Disappointingly, neither of these things
helped.
I imagine that including patterns would have to use prior knowledge from
game records (or wherever). Maybe they should not look like input