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 | -- | tactical library |
17 +------+ +--------+ +---------+ +------------------+
23 * "core" takes care of the program's lifetime, GTP interface and basic
24 fast Go board implementation
26 pachi.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 network.[ch] Network interface (useful for distributed engine)
32 timeinfo.[ch] Time-keeping information
33 stone.[ch] one board point coloring definition
34 move.[ch] one board move definition
35 board.[ch] board definition and basic interface
37 * "aux library" provides extra functions like static tactical evaluation
38 and pattern matching; it is somewhat interwound with "core" component
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
46 * "tactical library" provides extended interfaces for the go board,
47 most important non-trivial tactical information
51 * "engine" receives notifications about opponent moves and is asked
52 to generate a move to play on given board
54 engine.h abstract engine interface
55 random/ example "random move generator" engine
56 replay/ example "playout move generator" engine
57 montecarlo/ simple treeless Monte Carlo engine, quite bitrotten
58 uct/ the main UCT-player engine, see below
59 distributed/ "meta-engine" for distributed play by orchestrating
60 several UCT engines on different computers
61 joseki/ auxiliary engine for generating joseki patterns
63 * "playout" policy is asked to generate moves to play during the Monte Carlo
64 simulations, and to provide rough evaluation of moves feasibility for
67 playout.[ch] abstract playout policy interface,
68 Monte Carlo simulation execution
69 playout/light uniformly random playout policy
70 playout/moggy rule-based "Mogo-like" playout policy
72 * Also, several ways of testing Pachi are provided:
74 t-unit/ interface for writing unit-tests for specific
75 functionality, mainly tactics
76 t-play/ interface for testing performance by playing games
77 against a fixed opponent (e.g. GNUGo)
83 The UCT engine (the proper name should be MCTS as it does not have that
84 much common with classic UCT now) has non-trivial structure by itself:
86 +-------------+ +-----+ +-------------------+
87 | node policy | -- | UCT | --- | node prior-hinter |
88 +-------------+ +-----+ +-------------------+
94 * "UCT" is the core of the engine
96 uct.[ch] engine initialization, public interface
97 internal.h internal state and data structures
98 tree.[ch] minimax move tree with success statistics
99 walk.[ch] filling the tree by walking it many times
100 and running MC simulations from leaves
101 slave.[ch] engine interface for the distributed engine
103 * "node prior-hinter" assigns newly created nodes preliminary success
104 statistics ("prior values") to focus the search better
106 prior.[ch] variety of methods for setting the priors
108 * "node policy" mainly chooses the current node's child to descend
109 through during the tree walk, based on the already recorded statistics;
110 it must balance exploration and exploitation well during the selection
112 policy/ucb1 the old-school original simple policy
113 policy/ucb1amaf the AMAF/RAVE-based policy gathering statistics rapidly
115 * "dynkomi driver" dynamically determines self-imposed extra virtual komi
123 The infrastructure is optimized for speed to make it well suited
124 for bruteforce engines, however tradeoffs are made to make it useful
125 for heavier MonteCarlo playouts as well (e.g. real liberties are
126 tracked instead of pseudoliberties). If you are looking for raw
127 light playout speed, libEGO is better choice.
129 In general, arbitrary board sizes are supported; however, board sizes
130 smaller than 9x9 have not been tested and board sizes larger than 25x25
131 are not supported by GTP - also, in theory some buffer overflows might
132 happen with board sizes larger than 19x19. The engine parameters are
133 primarily tuned for 19x19 play.
138 While the Pachi engines generally play according to Chinese rules,
139 internally, Pachi uses Tromp-Taylor rules because they are simple,
140 fast and universal; they are very close to the New Zealand rules.
141 That means, it simply counts the number of stones and one-point eyes
142 of each color on the board, plus komi and handicap correction.
144 Tromp-Taylor rules also mean that multi-stone suicide is allowed! If you
145 do not like that (basically if you want to pretend it plays according
146 to Chinese rules), you need to rule that out in your engine, currently.
147 The provided engines DO avoid multi-stone suicide, though it is allowed
148 in the playouts for performance reasons (perhaps we should re-visit that
149 decision in light of heavy playouts).
151 Tromp-Taylor rules have positional superko; the board implementation
152 will set a flag if it is violated, but play the move anyway. You need
153 to enforce the superko rule in your engine.
159 ...is a very sad hack. ENSURE that only trusted parties talk to Pachi's
160 GTP interface, as it is totally non-resilient to any kind of overflow
161 or bad input attacks and allowing arbitrary input to be entered within
162 is a major security hole. Yes, this needs to be cleaned up. Also, currently
163 engines cannot plug in their own commands and there is no GoGui interface.
165 Pachi supports only few GTP commands now. Most importantly, it does not
166 support the undo command. The final_status_list command requires engine
173 The UCT engine allows external plugins to be loaded and provide external
174 knowledge for various heuristics - currently just biasing the MCTS priors,
175 but more can be added easily. The plugins should follow the API in
176 <uct/plugin.h> - see that file for details.
182 Joseki patterns can be generated from variations of a SGF file. Josekis
183 are matched per-quadrant, no other move must be within the quadrant. See
184 joseki/README for details on usage of the auxiliary engine and a full
185 recipe for making Pachi play according to joseki sequences extracted from
192 The UCT engine can "pre-read" the starting board position and
193 dump the core of the built tree to a file, loading it later. This is
194 called a 'tbook' (as in "tree book") and can be generated using the
195 tools/gentbook.sh script. The newly generated file is automatically
196 used by the UCT engine when found.
198 Alternatively, there is a support for directly used opening book
199 (so-called fbook, a.k.a. "forced book" or "fuseki book"). The book
200 is stored in a text file in Fuego-compatible format and can be loaded
201 using the ./pachi -f parameter. A naive way to build such a book
202 based on shell-based, twogtp-based UCT is available through the
203 tools/autobook/ framework.
209 This is a mostly unique idea in Pachi, currently in development;
210 it does not work too well yet, but Pasky has great hopes for it in
213 A local tree is a tree comprising of (CFG-topologically) local
214 sequences, built in parallel with the real game tree; there is
215 a separate tree for black-first and white-first sequences. E.g if
216 the game tree (on empty board) consists of D4-C6-D6-Q16-O17-D7,
217 the coresponding local tree sequences are (black) D4-C6-D6, (white)
218 Q16-O17 and (white) D7. The simulation result is then recorded as
219 usual in the game tree, but also in the corresponding local trees.
221 The goal of this is to dynamically build a cache of various
222 resolutions of local situations. This is to overcome cases where
223 the proper resolution of a given situation is easily found in
224 the tree, but improper heuristical bisaes keep muddying the
227 Exactly what kind of values to store in the local tree nodes is
228 still an open question. The original approach has been to simply
229 use simulation results directly or biased on base "value temperature";
230 now, we are experimenting with survival rate of the "base stone"
231 of the branch (the first move in the branch sequence). We also intend
232 to integrate criticality information with local tree data.
234 The local trees are descended in parallel with the main tree.
235 The values stored in local trees are added to the RAVE term
236 from the main tree during node selection. We also wish to use
237 the information from local trees in the simulations, but this is
238 still subject to research.
240 To enable local trees usage, pass local_tree=1 to the UCT engine.
241 There are many further knobs to twiddle, too.