UCT setup_state(): Allow first move to be white, just emit a warning

[pachi/nmclean.git] / HACKING

This is brief developer-oriented overview in Pachi structure.

Pachi is completely Go-specific (c.f. Fuego; though e.g. atari go support
should be easy to add), but fairly modular. It has been built with focus
on MonteCarlo-based play, but it can in principle be used for other
play engines as well.


Basic architecture
==================

Pachi consists of the following components:


  +------+    +--------+    +---------+    +------------------+
  | core | -- | engine | -- | playout | -- | tactical library |
  +------+    +--------+    +---------+    +------------------+
                     |        |
                  +-------------+
                  | aux library |
                  +-------------+

* "core" takes care of the program's lifetime, GTP interface and basic
  fast Go board implementation

        pachi.c         global initialization and the main loop
        version.h       current version information
        debug.h         debugging infrastructure
        random.[ch]     fast random number generator
        gtp.[ch]        GTP protocol interface
        network.[ch]    Network interface (useful for distributed engine)
        timeinfo.[ch]   Time-keeping information
        stone.[ch]      one board point coloring definition
        move.[ch]       one board move definition
        board.[ch]      board definition and basic interface

* "aux library" provides extra functions like static tactical evaluation
  and pattern matching; it is somewhat interwound with "core" component

        mq.h            "move queue" data structure
        stats.h         "move statistics" data structure
        probdist.[ch]   "probability distribution" data structure
        ownermap.[ch]   simulation-based finalpos. "owner map" data structure
        pattern3.[ch]   fast 3x3 spatial pattern matcher
        pattern.[ch]    general multi-feature pattern matcher

* "tactical library" provides extended interfaces for the go board,
  most important non-trivial tactical information

        tactics/

* "engine" receives notifications about opponent moves and is asked
  to generate a move to play on given board

        engine.h        abstract engine interface
        random/         example "random move generator" engine
        replay/         example "playout move generator" engine
        montecarlo/     simple treeless Monte Carlo engine, quite bitrotten
        uct/            the main UCT-player engine, see below
        distributed/    "meta-engine" for distributed play by orchestrating
                                several UCT engines on different computers
        patternscan/    auxiliary engine for harvesting patterns from
                                existing games
        joseki/         auxiliary engine for generating joseki patterns

* "playout" policy is asked to generate moves to play during the Monte Carlo
  simulations, and to provide rough evaluation of moves feasibility for
  the engine

        playout.[ch]    abstract playout policy interface,
                                Monte Carlo simulation execution
        playout/light   uniformly random playout policy
        playout/moggy   rule-based "Mogo-like" playout policy

* Also, several ways of testing Pachi are provided:

        t-unit/         interface for writing unit-tests for specific
                                functionality, mainly tactics
        t-play/         interface for testing performance by playing games
                                against a fixed opponent (e.g. GNUGo)


UCT architecture
================

The UCT engine (the proper name should be MCTS as it does not have that
much common with classic UCT now) has non-trivial structure by itself:

  +-------------+    +-----+     +-------------------+
  | node policy | -- | UCT | --- | node prior-hinter |
  +-------------+    +-----+     +-------------------+
                        |           |
                   +---------+      |
                   | playout | -----'
                   +---------+

* "UCT" is the core of the engine

        uct.[ch]        engine initialization, public interface
        internal.h      internal state and data structures
        tree.[ch]       minimax move tree with success statistics
        walk.[ch]       filling the tree by walking it many times
                                and running MC simulations from leaves
        slave.[ch]      engine interface for the distributed engine

* "node prior-hinter" assigns newly created nodes preliminary success
  statistics ("prior values") to focus the search better

        prior.[ch]      variety of methods for setting the priors

* "node policy" mainly chooses the current node's child to descend
  through during the tree walk, based on the already recorded statistics;
  it must balance exploration and exploitation well during the selection

        policy/ucb1     the old-school original simple policy
        policy/ucb1amaf the AMAF/RAVE-based policy gathering statistics rapidly

* "dynkomi driver" dynamically determines self-imposed extra virtual komi

        dynkomi.[ch]


Board Implementation
====================

The infrastructure is optimized for speed to make it well suited
for bruteforce engines, however tradeoffs are made to make it useful
for heavier MonteCarlo playouts as well (e.g. real liberties are
tracked instead of pseudoliberties). If you are looking for raw
light playout speed, libEGO is better choice.

In general, arbitrary board sizes are supported; however, board sizes
smaller than 9x9 have not been tested and board sizes larger than 25x25
are not supported by GTP - also, in theory some buffer overflows might
happen with board sizes larger than 19x19. The engine parameters are
primarily tuned for 19x19 play.

Ruleset
-------

While the Pachi engines generally play according to Chinese rules,
internally, Pachi uses Tromp-Taylor rules because they are simple,
fast and universal; they are very close to the New Zealand rules.
That means, it simply counts the number of stones and one-point eyes
of each color on the board, plus komi and handicap correction.

Tromp-Taylor rules also mean that multi-stone suicide is allowed! If you
do not like that (basically if you want to pretend it plays according
to Chinese rules), you need to rule that out in your engine, currently.
The provided engines DO avoid multi-stone suicide, though it is allowed
in the playouts for performance reasons (perhaps we should re-visit that
decision in light of heavy playouts).

Tromp-Taylor rules have positional superko; the board implementation
will set a flag if it is violated, but play the move anyway. You need
to enforce the superko rule in your engine.


GTP Implementation
==================

...is a very sad hack. ENSURE that only trusted parties talk to Pachi's
GTP interface, as it is totally non-resilient to any kind of overflow
or bad input attacks and allowing arbitrary input to be entered within
is a major security hole. Yes, this needs to be cleaned up. Also, currently
engines cannot plug in their own commands and there is no GoGui interface.

Pachi supports only few GTP commands now. Most importantly, it does not
support the undo command.  The final_status_list command requires engine
support.


General Pattern Matcher
=======================

Pachi has in-development general pattern matcher that can match various
sets of features (spatial and others), inspired by the CrazyStone and
van der Werf - de Groot - Stern pattern model. Please see pattern.h
for detailed description of the pattern concept and recognized features.

To harvest patterns, use 'pachi -e patternscan' (see patternscan/patternscan.c
for available options).  The output of the pattern scanner are two data
structures: The matched patterns

        (feature1:payload feature2:payload ...)

and spatial dictionary. "Spatial" feature represents a particular
configuration of stones in a circle around the move-to-play; each
configuration has its own record in the dictionary and the spatial
feature references only the id in the dictionary; so you need to keep
both the patterns and the "patterns.spat" file.  Normally, 'patternscan'
will only match already existing dictionary entries, but you
can pass it "gen_spat_dict" so that it appends all newly found spatial
features to the dictionary - use "spat_threshold" to limit recording
only to frequently occuring spatial features; to start the dictionary
from scratch, simply remove any existing "patterns.spat" file.

There are few pre-made scripts to make the initialization of the pattern
matcher easy:

* tools/pattern_byplayer.sh: Sorts out patterns from given SGF collection
  by player names, one player per file in a dedicated directory. This is
  useful if you want to use the patterns to e.g. recognize games of a
  player by characteristic patterns. Spatial dictionary is autogenerated
  in full.

* tools/pattern_spatial_gen.sh: Initializes spatial dictionary by spatial
  features found at least N times in given SGF collection.  This is useful
  for further gathering of general pattern statistics while keeping
  the amount of spatial features manageable.

* tools/pattern_spatial_show.pl ID: Shows spatial pattern of given id
  in 2D plane.

* tools/pattern_bayes_gen.sh: Generates a Bayesian probability table
  for each pattern in a SGF collection. The computed number is the empiric
  probability of playing a given pattern when it is available on board.

* tools/pattern_bayes_merge.sh: Merges Bayesian probability tables
  of patterns. This is useful for parallel pattern mining - run multiple
  pattern_bayes_gen (perhaps remove the counts >= 2 test) and then merge
  the generated tables.

* tools/pattern_getdrops.pl: Processes game logs to determine moves that
  lead to large drop in Pachi's evaluation and extracts patterns that
  represent these moves. This might allow Pachi to "learn from its past
  mistakes" and recognize these moves in the future. A full pipeline
  for this might be:

        cat kgsGtp.log | tools/kgslog2gtp.pl |
                tools/pattern_getdrops.pl | tools/pattern_bayes_gen.sh - |
                tools/pattern_bayes_merge.sh patterns.prob - >patterns.prob2


Plugin API
==========

The UCT engine allows external plugins to be loaded and provide external
knowledge for various heuristics - currently just biasing the MCTS priors,
but more can be added easily. The plugins should follow the API in
<uct/plugin.h> - see that file for details.


Joseki Patterns
===============

Joseki patterns can be generated from variations of a SGF file. Josekis
are matched per-quadrant, no other move must be within the quadrant. See
joseki/README for details on usage of the auxiliary engine and a full
recipe for making Pachi play according to joseki sequences extracted from
the Kogo dictionary.


Opening Book
============

The UCT engine can "pre-read" the starting board position and
dump the core of the built tree to a file, loading it later. This is
called a 'tbook' (as in "tree book") and can be generated using the
tools/gentbook.sh script. The newly generated file is automatically
used by the UCT engine when found.

Alternatively, there is a support for directly used opening book
(so-called fbook, a.k.a. "forced book" or "fuseki book"). The book
is stored in a text file in Fuego-compatible format and can be loaded
using the ./pachi -f parameter. A naive way to build such a book
based on shell-based, twogtp-based UCT is available through the
tools/autobook/ framework.


Local Trees
===========

This is a mostly unique idea in Pachi, currently in development;
it does not work too well yet, but Pasky has great hopes for it in
the future.

A local tree is a tree comprising of (CFG-topologically) local
sequences, built in parallel with the real game tree; there is
a separate tree for black-first and white-first sequences. E.g if
the game tree (on empty board) consists of D4-C6-D6-Q16-O17-D7,
the coresponding local tree sequences are (black) D4-C6-D6, (white)
Q16-O17 and (white) D7. The simulation result is then recorded as
usual in the game tree, but also in the corresponding local trees.

The goal of this is to dynamically build a cache of various
resolutions of local situations. This is to overcome cases where
the proper resolution of a given situation is easily found in
the tree, but improper heuristical bisaes keep muddying the
evaluation down.

Exactly what kind of values to store in the local tree nodes is
still an open question. The original approach has been to simply
use simulation results directly or biased on base "value temperature";
now, we are experimenting with survival rate of the "base stone"
of the branch (the first move in the branch sequence). We also intend
to integrate criticality information with local tree data.

The local trees are descended in parallel with the main tree.
The values stored in local trees are added to the RAVE term
from the main tree during node selection. We also wish to use
the information from local trees in the simulations, but this is
still subject to research.

To enable local trees usage, pass local_tree=1 to the UCT engine.
There are many further knobs to twiddle, too.