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 timeinfo.[ch] Time-keeping information
32 stone.[ch] one board point coloring definition
33 move.[ch] one board move definition
34 board.[ch] board definition and basic interface
36 * "aux library" provides extra functions like static tactical evaluation
37 and pattern matching; it is somewhat interwound with "core" component
39 tactics.[ch] extended interfaces for the go board
40 mq.h "move queue" data structure
41 stats.h "move statistics" data structure
42 probdist.[ch] "probability distribution" data structure
43 ownermap.[ch] simulation-based finalpos. "owner map" data structure
44 pattern3.[ch] fast 3x3 spatial pattern matcher
45 pattern.[ch] general multi-feature pattern matcher
46 network.[ch] Network interface (useful for distributed engine)
48 * "engine" receives notifications about opponent moves and is asked
49 to generate a move to play on given board
51 engine.h abstract engine interface
52 random/ example "random move generator" engine
53 replay/ example "playout move generator" engine
54 montecarlo/ simple treeless Monte Carlo engine, quite bitrotten
55 uct/ the main UCT-player engine, see below
56 patternscan/ auxiliary engine for harvesting patterns from
58 distributed/ "meta-engine" for distributed play by orchestrating
59 several UCT engines on different computers
61 * "playout" policy is asked to generate moves to play during the Monte Carlo
62 simulations, and to provide rough evaluation of moves feasibility for
65 playout.[ch] abstract playout policy interface,
66 Monte Carlo simulation execution
67 playout/light uniformly random playout policy
68 playout/moggy rule-based "Mogo-like" playout policy
69 playout/elo probdist-based "CrazyStone-like" playout policy
71 * Also, several ways of testing Pachi are provided:
73 t-unit/ interface for writing unit-tests for specific
74 functionality, mainly tactics
75 t-play/ interface for testing performance by playing games
76 against a fixed opponent (e.g. GNUGo)
82 The UCT engine has non-trivial structure by itself:
84 +-------------+ +-----+ +-------------------+
85 | node policy | -- | UCT | --- | node prior-hinter |
86 +-------------+ +-----+ +-------------------+
92 * "UCT" is the core of the engine
94 uct.[ch] engine initialization, public interface
95 internal.h internal state and data structures
96 tree.[ch] minimax move tree with success statistics
97 walk.[ch] filling the tree by walking it many times
98 and running MC simulations from leaves
100 * "node prior-hinter" assigns newly created nodes preliminary success
101 statistics ("prior values") to focus the search better
103 prior.[ch] variety of methods for setting the priors
105 * "node policy" mainly chooses the current node's child to descend
106 through during the tree walk, based on the already recorded statistics;
107 it must balance exploration and exploitation well during the selection
109 policy/ucb1 the old-school original simple policy
110 policy/ucb1amaf the AMAF/RAVE-based policy gathering statistics rapidly
116 The infrastructure is optimized for speed to make it well suited
117 for bruteforce engines, however tradeoffs are made to make it useful
118 for heavier MonteCarlo playouts as well (e.g. real liberties are
119 tracked instead of pseudoliberties). If you are looking for raw
120 light playout speed, libEGO is better choice.
125 While the Pachi engines generally play according to Chinese rules,
126 internally, Pachi uses Tromp-Taylor rules because they are simple,
127 fast and universal; they are very close to the New Zealand rules.
128 That means, it simply counts the number of stones and one-point eyes
129 of each color on the board, plus komi and handicap correction.
131 Tromp-Taylor rules also mean that multi-stone suicide is allowed! If you
132 do not like that (basically if you want to pretend it plays according
133 to Chinese rules), you need to rule that out in your engine, currently.
134 The provided engines DO avoid multi-stone suicide (but the UCT engine
135 will never play it itself).
137 Tromp-Taylor rules have positional superko; the board implementation
138 will set a flag if it is violated, but play the move anyway. You need
139 to enforce the superko rule in your engine.
145 ...is a very sad hack. ENSURE that only trusted parties talk to Pachi's
146 GTP interface, as it is totally non-resilient to any kind of overflow
147 or bad input attacks and allowing arbitrary input to be entered within
148 is a major security hole. Yes, this needs to be cleaned up. Also, currently
149 engines cannot plug in their own commands and there is no GoGui interface.
151 Pachi supports only few GTP commands now. Most importantly, it does not
152 support the undo command. The final_status_list command requires engine
156 General Pattern Matcher
157 =======================
159 Pachi has in-development general pattern matcher that can match various
160 sets of features (spatial and others), inspired by the CrazyStone pattern
161 model. Please see pattern.h for detailed description of the pattern concept
162 and recognized features.
164 To harvest patterns, use 'zzgo -e patternscan' (see patternscan/patternscan.c
165 for available options). The output of the pattern scanner are two data
166 structures: The matched patterns
168 (feature1:payload feature2:payload ...)
170 and spatial dictionary. "Spatial" feature represents a particular
171 configuration of stones in a circle around the move-to-play; each
172 configuration has its own record in the dictionary and the spatial
173 feature references only the id in the dictionary; so you need to keep
174 both the patterns and the "patterns.spat" file. Normally, 'patternscan'
175 will only match already existing dictionary entries, but you
176 can pass it "gen_spat_dict" so that it appends all newly found spatial
177 features to the dictionary - use "spat_threshold" to limit recording
178 only to frequently occuring spatial features; to start the dictionary
179 from scratch, simply remove any existing "patterns.spat" file.
181 There are few pre-made scripts to make the initialization of the pattern
184 * pattern_byplayer.sh: Sorts out patterns from given SGF collection by
185 player names, one player per file in a dedicated directory. This is
186 useful if you want to use the patterns to e.g. recognize games of a
187 player by characteristic patterns. Spatial dictionary is autogenerated
190 * pattern_spatial_gen.sh: Initializes spatial dictionary by spatial features
191 found at least N times in given SGF collection. This is useful for
192 further gathering of general pattern statistics while keeping the amount
193 of spatial features manageable.
195 * pattern_spatial_show.pl ID: Shows spatial pattern of given id in 2D plane.
197 * pattern_mm.sh: Combines patternsacn engine and the MM tool (see below),
198 producing gamma values for harvested patterns.
200 Minorization-majorization (CrazyStone patterns)
201 -----------------------------------------------
203 The pattern harvester can be used together with the MM tool by Remi Coulom:
205 http://remi.coulom.free.fr/Amsterdam2007/mm.tar.bz2
207 This tool will compute relative strength of individual features for teaming
208 them up and using the outcoming probability distribution for generating moves.
209 There is a script that will take you from SGF game collection to gamma values
210 in single shot - "pattern_mm.sh".
212 The resulting "patterns.gamma" file contains mapping from feature instances
213 to gamma floats, representing the features strength; note that it is totally
214 meaningless without the accompanying "patterns.spat" file generated by the
215 pattern_gather script. To make Pachi use the gamma values for tree bias and
216 in MC playouts, use the "elo" playout policy - but note that it's still in