[gnugo-devel] Tunable Monte Carlo patterns

[Gunnar Farnebäck]

Here's an Easter surprise. I have added a patch to CVS which makes big
changes to the Monte Carlo code.

First it adds incremental updates of local 3x3 neighborhood, suicide
status, self-atari status, and number of stones captured, for each
move.

Second it makes the random playout move generation distributed
strictly proportional to move values computed by table lookup from a
local context consisting of 3x3 neighborhood, opponent suicide status,
own and opponent self-atari status, number of stones captured by own
and opponent move, and closeness to the previous move. Let's call this
local context simply "a pattern" and the table "pattern values" or
simply "patterns".

Third it makes the pattern values tunable. More about this below.

The net result is that the Monte Carlo code has become about 30%
faster than before and much more flexible.

-----------

Like before, the Monte Carlo mode is only supported for 9x9. It is
enabled by the "--monte-carlo" command line option and the speed can
be varied by the "--mc-games-per-level <n>" option.

There are three new options. To list available compiled in pattern
values, use "--mc-list-patterns", which currently gives the result

* montegnu_classic (default)
* mogo_classic
* uniform

The first of these is an approximation of the previous random move
generation algorithm. The mogo_classic pattern values is an
approximation of the simulation policy used by early versions of MoGo,
as published in the report "Modification of UCT with Patterns in
Monte-Carlo Go", RR-6062, by Sylvain Gelly, Yizao Wang, Rémi Munos,
and Olivier Teytaud. The uniform pattern values is the so called
"light" playout which chooses uniformly between all legal moves except
single point proper eyes.

To choose one of these, use e.g. "--mc-patterns mogo_classic".

However, if you're not satisfied with these you can also tune your own
pattern values with a pattern database file and load it at runtime
with e.g. "--mc-load-patterns my_mc_patterns.db". The rest of this
message will show how.

-----------

Let's start with the uniform pattern values. Those are defined by the
file patterns/mc_uniform.db, which looks like this:

oOo
O*O
oO?

:0

oOo
O*O
---

:0

|Oo
|*O
+--

:0

Patterns are always exactly 3x3 in size with the move at the center
point. The symbols are the usual for GNU Go pattern databases:

* move
O own stone (i.e. the same color as the color to move)
o own stone or empty
X opponent stone
x opponent stone or empty
? own stone, opponent stone, or empty
| vertical edge
- horizontal edge
+ corner

There's also a new symbol:

% own stone, opponent stone, empty, or edge

After the pattern comes a line starting with a colon. In all these
patterns it says that the pattern has a move value of 0, i.e. must not
be played. Unmatched patterns have a default value of 1. When all move
values are zero for both players, the playout will stop. Including the
three patterns above is important because otherwise the playouts would
be likely to go on indefinitely, or as it actually happens be
terminated at a hard-coded limit of 600 moves. Also place these
patterns at the top of the database because when multiple patterns
match, the first one is used, regardless of the values.

When using only these patterns you will probably notice that it plays
rather heavy, trying hard to be solidly connected. This is because
uniform playouts are badly biased with a high probability of non-solid
connections being cut apart. To counter this you could try a pattern
like

?X?
O*O
x.?

:20,near

to increase the probability that the one-point jump is reinforced when
threatened. Here we added the property "near", which means that the
pattern only applies if the previous move was played "near" this move.
Primarily "near" means within the surrounding 3x3 neighborhood but it
also includes certain cases of liberties of low-liberty strings
adjacent to the previous move, e.g. the move to extend out of an atari
created by the previous move. You have to read the source to find out
the exact rules for nearness.

We could also be even more specific and say

?X?
O*O
x.?

:20,near,osafe,xsafe

to exclude the cases where this move is a self atari (osafe) or would
be a self-atari for the opponent (xsafe).

It may also be interesting to see the effect of capturing stones. A
catch-all pattern for captures would be

?X%
?*%
%%%

:10,ocap1,osafe
:20,ocap2
:30,ocap3

where we have used multiple colon lines to specify different move
values depending on the number of captured stones; value 10 for a
single captured stone, value 20 for two captured stones, and value 30
for three or more captured stones. Here we also excluded self-atari
moves in the case of 1 captured stone in order to avoid getting stuck
in triple-ko in the playouts (there's no superko detection in the
playouts).

The full set of pattern properties is as follows:
near        The move is "near" the previous move.
far         The move is not "near" the previous move.
osafe       The move is not a self-atari.
ounsafe     The move is a self-atari.
xsafe       The move would not be a self-atari for the opponent.
xunsafe     The move would be a self-atari for the opponent.
xsuicide    The move would be suicide for the opponent
xnosuicide  The move would not be suicide for the opponent.
ocap0       The move captures zero stones.
ocap1       The move captures one stone.
ocap2       The move captures two stones.
ocap3       The move captures three or more stones.
ocap1+      The move captures one or more stones.
ocap1-      The move captures at most one stones.
ocap2+      The move captures two or more stones.
ocap2-      The move captures at most two stones.
xcap0       An opponent move would capture zero stones.
xcap1       An opponent move would capture one stone.
xcap2       An opponent move would capture two stones.
xcap3       An opponent move would capture three or more stones.
xcap1+      An opponent move would capture one or more stones.
xcap1-      An opponent move would capture at most one stones.
xcap2+      An opponent move would capture two or more stones.
xcap2-      An opponent move would capture at most two stones.

These can be combined arbitrarily but all must be satisfied for the
pattern to take effect. If contradictory properties are combined, the
pattern will never match.

Final comments:
* Move values are unsigned 32-bit integers. To avoid overflow in
  computations it is highly recommended to keep the values below
  10000000 or so.

* There is no speed penalty for having lots of patterns in the
  database. The average time per move is approximately constant
  (slightly dependent on how often stones are captured or become low
  on liberties) and the time per game mostly depends on the average
  game length.

* For more complex pattern databases, see
  patterns/mc_montegnu_classic.db and patterns/mc_mogo_classic.db.

-----------

I don't know the relative strength of the three builtin pattern
databases. Please try them out on CGOS if you have some spare computer
power.

Nobody really knows how to tune the random playouts to get as strong
engine as possible. Please play with this and report any interesting
findings, especially if you're able to make it substantially stronger
than the montegnu_classic patterns.

Have fun!

/Gunnar