1 This is brief developer-oriented overview in Pachi structure.
3 Pachi is completely Go-specific (c.f. Fuego; though e.g. atari go support
4 should be easy to add), but fairly modular. It has been built with focus
5 on MonteCarlo-based play, but it can in principle be used for other
12 Pachi consists of the following components:
15 +------+ +--------+ +---------+
16 | core | -- | engine | -- | playout |
17 +------+ +--------+ +---------+
23 * "core" takes care of the program's lifetime, GTP interface and basic
24 fast Go board implementation
26 zzgo.c global initialization and the main loop
27 version.h current version information
28 debug.h debugging infrastructure
29 random.[ch] fast random number generator
30 gtp.[ch] GTP protocol interface
31 stone.[ch] one board point coloring definition
32 move.[ch] one board move definition
33 board.[ch] board definition and basic interface
35 * "aux library" provides extra functions like static tactical evaluation
36 and pattern matching; it is somewhat interwound with "core" component
38 tactics.[ch] extended interfaces for the go board
39 mq.h "move queue" data structure
40 stats.h "move statistics" data structure
41 probdist.[ch] "probability distribution" data structure
42 ownermap.[ch] simulation-based finalpos. "owner map" data structure
43 pattern3.[ch] fast 3x3 spatial pattern matcher
44 pattern.[ch] general multi-feature pattern matcher
46 * "engine" receives notifications about opponent moves and is asked
47 to generate a move to play on given board
49 engine.h abstract engine interface
50 random/ example "random move generator" engine
51 replay/ example "playout move generator" engine
52 montecarlo/ simple treeless Monte Carlo engine, quite bitrotten
53 uct/ the main UCT-player engine, see below
54 patternscan/ auxiliary engine for harvesting patterns from
57 * "playout" policy is asked to generate moves to play during the Monte Carlo
58 simulations, and to provide rough evaluation of moves feasibility for
61 playout.[ch] abstract playout policy interface,
62 Monte Carlo simulation execution
63 playout/light uniformly random playout policy
64 playout/moggy rule-based "Mogo-like" playout policy
65 playout/elo probdist-based "CrazyStone-like" playout policy
67 * Also, several ways of testing Pachi are provided:
69 t-unit/ interface for writing unit-tests for specific
70 functionality, mainly tactics
71 t-play/ interface for testing performance by playing games
72 against a fixed opponent (e.g. GNUGo)
78 The UCT engine has non-trivial structure by itself:
80 +-------------+ +-----+ +-------------------+
81 | node policy | -- | UCT | --- | node prior-hinter |
82 +-------------+ +-----+ +-------------------+
88 * "UCT" is the core of the engine
90 uct.[ch] engine initialization, public interface
91 internal.h internal state and data structures
92 tree.[ch] minimax move tree with success statistics
93 walk.[ch] filling the tree by walking it many times
94 and running MC simulations from leaves
96 * "node prior-hinter" assigns newly created nodes preliminary success
97 statistics ("prior values") to focus the search better
99 prior.[ch] variety of methods for setting the priors
101 * "node policy" mainly chooses the current node's child to descend
102 through during the tree walk, based on the already recorded statistics;
103 it must balance exploration and exploitation well during the selection
105 policy/ucb1 the old-school original simple policy
106 policy/ucb1amaf the AMAF/RAVE-based policy gathering statistics rapidly
112 The infrastructure is optimized for speed to make it well suited
113 for bruteforce engines, however tradeoffs are made to make it useful
114 for heavier MonteCarlo playouts as well (e.g. real liberties are
115 tracked instead of pseudoliberties). If you are looking for raw
116 light playout speed, libEGO is better choice.
121 While the Pachi engines generally play according to Chinese rules,
122 internally, Pachi uses Tromp-Taylor rules because they are simple,
123 fast and universal; they are very close to the New Zealand rules.
124 That means, it simply counts the number of stones and one-point eyes
125 of each color on the board, plus komi and handicap correction.
127 Tromp-Taylor rules also mean that multi-stone suicide is allowed! If you
128 do not like that (basically if you want to pretend it plays according
129 to Chinese rules), you need to rule that out in your engine, currently.
130 The provided engines DO avoid multi-stone suicide (but the UCT engine
131 will never play it itself).
133 Tromp-Taylor rules have positional superko; the board implementation
134 will set a flag if it is violated, but play the move anyway. You need
135 to enforce the superko rule in your engine.
141 ...is a very sad hack. ENSURE that only trusted parties talk to Pachi's
142 GTP interface, as it is totally non-resilient to any kind of overflow
143 or bad input attacks and allowing arbitrary input to be entered within
144 is a major security hole. Yes, this needs to be cleaned up. Also, currently
145 engines cannot plug in their own commands and there is no GoGui interface.
147 Pachi supports only few GTP commands now. Most importantly, it does not
148 support the undo command and it does not support time-keeping.
149 The final_status_list command requires engine support.
152 General Pattern Matcher
153 =======================
155 Pachi has in-development general pattern matcher that can match various
156 sets of features (spatial and others), inspired by the CrazyStone pattern
157 model. Please see pattern.h for detailed description of the pattern concept
158 and recognized features.
160 To harvest patterns, use 'zzgo -e patternscan' (see patternscan/patternscan.c
161 for available options). The output of the pattern scanner are two data
162 structures: The matched patterns
164 (feature1:payload feature2:payload ...)
166 and spatial dictionary. "Spatial" feature represents a particular
167 configuration of stones in a circle around the move-to-play; each
168 configuration has its own record in the dictionary and the spatial
169 feature references only the id in the dictionary; so you need to keep
170 both the patterns and the "patterns.spat" file. Normally, 'patternscan'
171 will only match already existing dictionary entries, but you
172 can pass it "gen_spat_dict" so that it appends all newly found spatial
173 features to the dictionary - use "spat_threshold" to limit recording
174 only to frequently occuring spatial features; to start the dictionary
175 from scratch, simply remove any existing "patterns.spat" file.
177 There are few pre-made scripts to make the initialization of the pattern
180 * pattern_byplayer.sh: Sorts out patterns from given SGF collection by
181 player names, one player per file in a dedicated directory. This is
182 useful if you want to use the patterns to e.g. recognize games of a
183 player by characteristic patterns. Spatial dictionary is autogenerated
186 * pattern_spatial_gen.sh: Initializes spatial dictionary by spatial features
187 found at least N times in given SGF collection. This is useful for
188 further gathering of general pattern statistics while keeping the amount
189 of spatial features manageable.
191 * pattern_spatial_show.pl ID: Shows spatial pattern of given id in 2D plane.
193 * pattern_mm.sh: Combines our scripts and the MM tool (see below),
194 producing gamma values for harvested patterns.
196 These extra scripts are used internally by pattern_mm:
198 * pattern_enumerate.pl: Numbers all patterns and possible payloads and
199 produces numeric versions of patterns and a mapping file.
201 * pattern_mminput.pl: Takes output of pattern_enumerate.pl and converts
202 it to a format that can be directly pipelined to the MM tool (see below).
204 Minorization-majorization (CrazyStone patterns)
205 -----------------------------------------------
207 The pattern harvester can be used together with the MM tool by Remi Coulom:
209 http://remi.coulom.free.fr/Amsterdam2007/mm.tar.bz2
211 This tool will compute relative strength of individual features for teaming
212 them up and using the outcoming probability distribution for generating moves.
213 There is a script that will take you from SGF game collection to gamma values
214 in single shot - "pattern_mm.sh".
216 The resulting "patterns.gamma" file contains mapping from feature instances
217 to gamma floats, representing the features strength; note that it is totally
218 meaningless without the accompanying "patterns.spat" file generated by the
219 pattern_gather script. To make Pachi use the gamma values for tree bias and
220 in MC playouts, use the "elo" playout policy - but note that it's still in