This chapter is an overview of the GNU Go internals. Further
documentation of how any one module or routine works may be found in
later chapters or comments in the source files.
GNU Go starts by trying to understand the current board position as
good as possible. Using the information found in this first phase, and
using additional move generators, a list of candidate moves is generated.
Finally, each of the candidate moves is valued according to its territorial
value (including captures or life-and-death effects), and possible
strategical effects (such as strengthening a weak group).
Note that while GNU Go does, of course, do a lot of reading to analyze
possible captures, life and death of groups etc., it does not (yet) have
* Examining the Position:: Gathering Information
* Move Generators:: Selecting Candidate Moves
* Move Valuation:: Selecting the best Move
* Detailed Sequence of Events:: Outline of @code{genmove()}.
* Roadmap:: Description of the different files.
* Coding Styles:: Coding conventions.
* Navigating the Source:: Navigating the Source.
@node Examining the Position
@section Gathering Information
This is by far the most important phase in the move generation.
Misunderstanding life-and-death situations can cause gross mistakes.
Wrong territory estimates will lead to inaccurate move valuations.
Bad judgement of weaknesses of groups make strategic mistakes likely.
This information gathering is done by the function @code{examine_position()}.
It first calls @code{make_worms()}.
Its first steps are very simple: it identifies sets of directly connected
stones, called @dfn{worms}, and notes their sizes and their number of
Soon after comes the most important step of the worm analysis:
the tactical reading code (@pxref{Tactical Reading}) is called for every
out which worms can be captured directly, giving up as soon as a worm
can reach 5 liberties. If a worm can be captured, the engine of course
looks for moves defending against this capture. Also, a lot of effort
is made to find virtually all moves that achieve the capture or defense
After knowing which worms are tactically stable, we can make a first
picture of the balance of power across the board: the @ref{Influence}
code is called for the first time.
This is to aid the next step, the analysis of dragons. By a @dfn{dragon}
we mean a group of stones that cannot be disconnected.
Naturally the first step in the responsible function @code{make_dragons()}
is to identify these dragons, i.e. determine which worms cannot be
disconnected from each other. This is partly done by patterns, but
in most cases the specialized readconnect code
@comment FIXME: Put in cross-ref here once Connection is documented
is called. This module does a minimax search to determine whether two
given worms can be connected with, resp. disconnected from each other.
Then we compute various measures to determine how strong or weak any given
@item A crude estimate of the number of eyes is made.
@item The results of the influence computations is used to see which dragons
are adjacent to own territory or a moyo.
@item A guess is made for the potential to escape if the dragon got
For those dragons that are considered weak, a life and death analysis
is made (@pxref{The Owl Code}). If two dragons next to each other are found
that are both not alive, we try to resolve this situation with the semeai
For a more detailed reference of the worm and dragon analysis (and
explanations of the data structures used to store the information),
see @xref{Worms and Dragons}.
The influence code is then called second time to make a detailed analysis
of likely territory. Of course, the life-and-death status of dragons are
The territorial results of the influence module get corrected by the break-in
module. This specifically tries to analyze where an opponent could break
into an alleged territory, with sequences that would be too difficult to
see for the influence code.
Once we have found out all about the position it is time to generate
the best move. Moves are proposed by a number of different modules
called @dfn{move generators}. The move generators themselves
do not set the values of the moves, but enumerate justifications for
them, called @dfn{move reasons}. The valuation of the moves comes
last, after all moves and their reasons have been generated.
For a list and explanation of move reasons used in GNU Go, and how they
are evaluated, see @xref{Move Generation}.
There are a couple of move generators that only extract data found in
the previous phase, examining the position:
@item @code{worm_reasons()}
Moves that have been found to capture or defend a worm are proposed as
@item @code{owl_reasons()}
The status of every dragon, as it has been determined by the owl code
(@pxref{The Owl Code}) in the previous phase, is reviewed. If the status
is critical, the killing or defending move gets a corresponding move
@item @code{semeai_move_reasons()}
Similarly as @code{owl_reasons}, this function proposes moves relevant
@item @code{break_in_move_reasons()}
This suggests moves that have been found to break into opponent's territory
The following move generators do additional work:
Generate a move in the early fuseki, either in an empty corner of from
This is probably the most important move generator.
It finds patterns from @file{patterns/patterns.db},
@file{patterns/patterns2.db}, @file{patterns/fuseki.db}, and the joseki
files in the current position. Each pattern is matched in each
of the 8 possible orientations obtainable by rotation and
reflection. If the pattern matches, a so called "constraint"
may be tested which makes use of reading to determine if the
pattern should be used in the current situation. Such
constraints can make demands on number of liberties of
strings, life and death status, and reading out ladders,
etc. The patterns may call helper functions, which may
be hand coded (in @file{patterns/helpers.c}) or
The patterns can be of a number of different classes
with different goals. There are e.g. patterns which
try to attack or defend groups, patterns which try to
connect or cut groups, and patterns which simply try
to make good shape. (In addition to the large pattern
database called by @code{shapes()}, pattern matching
is used by other modules for different tasks throughout
the program. @xref{Patterns}, for a complete documentation
@item @code{combinations()}
See if there are any combination threats or atari sequences and either
propose them or defend against them.
@item @code{revise_thrashing_dragon()}
@findex revise_thrashing_dragon
This module does not directly propose move: If we are clearly ahead,
and the last move played by the opponent is part of a dead dragon, we
want to attack that dragon again to be on the safe side. This is done
be setting the status of this @dfn{thrashing dragon} to unkown and
repeating the shape move generation and move valution.
@item @code{endgame_shapes()}
If no move is found with a value greater than 6.0, this module matches a
set of extra patterns which are designed for the endgame. The endgame
patterns can be found in @file{patterns/endgame.db}.
@item @code{revise_semeai()}
If no move is found, this module changes the status of opponent groups
involved in a semeai from @code{DEAD} to @code{UNKNOWN}. After this,
genmove runs @code{shapes} and @code{endgame_shapes} again to see if a
@item @code{fill_liberty()}
Fill a common liberty. This is only used at the end
of the game. If necessary a backfilling or backcapturing
After the move generation modules have run, each proposed candidate
move goes through a detailed valuation by the function
@code{review_move_reasons}. This invokes some analysis to try to turn
up other move reasons that may have been missed.
The most important value of a move is its territorial effect.
@pxref{Influence and Territory} explains in detail how this is determined.
This value is modified for all move reasons that cannot be expressed
directly in terms of territory, such as combination attacks (where it
is not clear which of several strings will get captured), strategical
effects, connection moves, etc. A large set heuristics is necessary
here, e.g. to avoid duplication of such values. This is explained in
more detail in @ref{Valuation}.
@node Detailed Sequence of Events
@section Detailed Sequence of Events
First comes the sequence of events when
@code{examine_position()} is run from @code{genmove()}. This
@code{purge_persistent_caches()}
@code{compute_effective_sizes()}
@code{compute_unconditional_status()}
@code{find_worm_attacks_and_defenses()}:
for each attackable worm:
@code{change_attack()} to add the attack point
@code{find_attack_patterns()} to find a few more attacks
for each defensible worm:
@code{change_defense()} to add the defense point
@code{find_defense_patterns()} to find a few more defense moves
find additional attacks and defenses by testing all
find higher order liberties (for each worm)
find cutting stones (for each worm)
improve attacks and defenses: if capturing a string defends
another friendly string, or kills an unfriendly one, we
add points of defense or attack. Make repairs if adjacent
strings can both be attacked but not defended.
identify inessential worms (such as nakade stones)
@code{compute_worm_influence()}:
@code{find_influence_patterns()}
@code{segment_influence()}
@code{find_connections()}
@code{make_domains()} (determine eyeshapes)
@code{find_lunches()} (adjacent strings that can be captured)
@code{find_half_and_false_eyes()}
@code{eye_computations()}: Compute the value of each eye space.
Store its attack and defense point.
@code{analyze_false_eye_territory()}
for each dragon @code{compute_dragon_genus()}
for each dragon @code{compute_escape()} and set escape route data
@code{resegment_initial_influence()}
@code{compute_refined_dragon_weaknesses()} (called again after owl)
for each dragon @code{compute_crude_status()}
@code{find_neighbor_dragons()}
for each dragon compute surround status
for each weak dragon run @code{owl_attack()} and @code{owl_defend()}
to determine points of attack and defense
for each dragon compute dragon.status
for each thrashing dragon compute owl threats
for each dragon compute dragon.safety
@code{revise_inessentiality()}
for every semeai, run @code{owl_analyze_semeai()}
@code{find_moves_to_make_seki()}
@code{identify_thrashing_dragons()}
@code{compute_dragon_influence()}:
@code{compute_influence()}
@code{break_territories()} (@pxref{Break Ins})
@code{compute_refined_dragon_weaknesses()}
Now a summary of the sequence of events during the
move generation and selection phases of @code{genmove()}, which
take place after the information gathering phase has been completed:
@code{collect_move_reasons()}:
@code{worm_reasons()}: for each attack and defense point add a move reason
@code{semeai_reasons()}: for each dragon2.semeai point add a move reason
@code{owl_reasons()}: for each owl attack and defense point add a move reason
@code{break_in_reasons()}: for each breakin found add a move reason
@code{shapes()}: match patterns around the board (@pxref{Patterns Overview})
@code{combinations()}: look for moves with a double meaning and other tricks
@code{find_double_threats()}
@code{review_move_reasons()}
if ahead and there is a thrashing dragon, consider it
alive and reconsider the position
if no move found yet, revisit any semeai, change status of dead opponent
to alive, then run @code{shapes()} and @code{endgame_shapes()} again
if no move found yet, run @code{fill_liberty()}
The GNU Go engine is contained in two directories, @file{engine/} and
@file{patterns/}. Code related to the user interface, reading and
writing of Smart Game Format files, and testing are found in the
directories @file{interface/}, @file{sgf/}, and @file{regression/}. Code
borrowed from other GNU programs is contained in @file{utils/}. That
directory also includes some code developed within GNU Go which is not
go specific. Documentation is in @file{doc/}.
In this document we will describe some of the individual files comprising
the engine code in @file{engine/} and @file{patterns/}. In @file{interface/}
This is the Go Modem Protocol interface (courtesy of
William Shubert and others). This takes care of all the
details of exchanging setup and moves with Cgoban, or any
other driving program recognizing the Go Modem Protocol.
This contains @code{main()}. The @file{gnugo} target is
thus built in the @file{interface/} directory.
@subsection Files in @file{engine/}
In @file{engine/} there are the following files:
Contains algorithms which may be called at the end of the game to generate
moves that will generate moves to settle the position, if necessary playing
out a position to determine exactly the status of every group on the board,
which GNU Go can get wrong, particularly if there is a seki. This module is
the basis for the most accurate scoring algorithm available in GNU Go.
This file contains code for the maintenance of the board. For example
it contains the important function @code{trymove()} which tries a move
on the board, and @code{popgo()} which removes it by popping the move
stack. At the same time vital information such as the number of
liberties for each string and their location is updated incrementally.
Code to detect moves which can break into supposed territory and moves
@item @file{cache.c} and @file{cache.h}
As a means of speeding up reading, computed results are cached so that
they can be quickly reused if the same position is encountered through
e.g. another move ordering. This is implemented using a hash table.
@item @file{clock.c} and @file{clock.h}
Clock code, including code allowing GNU Go to automatically
adjust its level in order to avoid losing on time in tournaments.
@item @file{combination.c}
When something can (only) be captured through a series of ataris or
other threats we call this a combination attack. This file contains code
to find such attacks and moves to prevent them.
This contains @code{make_dragons()}. This function is executed before
the move-generating modules @code{shapes()} @code{semeai()} and the
other move generators but after @code{make_worms()}. It tries to connect
worms into dragons and collect important information about them, such as
how many liberties each has, whether (in GNU Go's opinion) the dragon
can be captured, if it lives, etc.
Code to find certain types of endgame moves.
Code to force filling of dame (backfilling if necessary)
Generates fuseki (opening) moves from a database. Also generates moves
This file contains @code{genmove()} and its supporting
routines, particularly @code{examine_position()}.
This contains the principal global variables used by GNU Go.
This file contains declarations forming the public interface to
@item @file{hash.c} and @file{hash.h}
Hashing code implementing Zobrist hashing. (@pxref{Hashing}) The code in
@file{hash.c} provides a way to hash board positions into compact descriptions
which can be efficiently compared. The caching code in @file{cache.c}
makes use of the board hashes when storing and retrieving read results.
@item @file{influence.c} and @file{influence.h}.
This code determines which regions of the board are under the
influence of either player.
Header file for the engine. The name ``liberty'' connotes
freedom (@pxref{Copying}).
This file contains the pattern matcher @code{matchpat()}, which looks for
patterns at a particular board location. The actual patterns are in
the @file{patterns/} directory. The function @code{matchpat()} is
called by every module which does pattern matching, notably @code{shapes}.
@item @file{move_reasons.c} and @file{move_reasons.h}
Code for keeping track of move reasons.
Supporting code for lists of moves.
This file contains the code to recognize eye shapes,
documented in @xref{Eyes}.
Code to fork off a second GNU Go process which can be used to simulate
reading with top level information (e.g. dragon partitioning) available.
This file does life and death reading. Move generation is pattern based
and the code in @file{optics.c} is used to evaluate the eyespaces for
vital moves and independent life. A dragon can also live by successfully
escaping. Semeai reading along the same principles is also implemented
@item @file{persistent.c}
Persistent cache which allows reuse of read results at a later move or
with additional stones outside an active area, which are those
intersections thought to affect the read result.
@item @file{printutils.c}
@item @file{readconnect.c} and @file{readconnect.h}
This file contains code to determine whether two strings can be
connected or disconnected.
This file contains code to determine whether any given
string can be attacked or defended. @xref{Tactical Reading},
This file contains @code{semeai()}, the module which detects dragons
in semeai. To determine the semeai results the semeai reading in
Code to generate sgf traces for various types of reading.
This file contains @code{shapes()}, the module called by @code{genmove()}
which tries to find moves which match a pattern (@pxref{Patterns}).
This file contains @code{showboard()}, which draws an ASCII
representation of the board, depicting dragons (stones
with same letter) and status (color). This was the
primary interface in GNU Go 1.2, but is now a debugging
Code to determine whether a dragon is surrounded and to find moves to
surround with or break out with.
An assortment of utilities, described in greater detail below.
@item @file{value_moves.c}
This file contains the code which assigns values to every move
after all the move reasons are generated. It also tries to generate
certain kinds of additional move reasons.
This file contains @code{make_worms()}, code which is run at the
beginning of each move cycle, before the code in @file{dragon.c}, to
determine the attributes of every string. These attributes are things
like liberties, wether the string can be captured (and how), etc
@subsection Files in @file{patterns/}
The directory @file{patterns/} contains files related to pattern matching.
Currently there are several types of patterns. A partial list:
@item move generation patterns in @file{patterns.db} and @file{patterns2.db}
@item move generation patterns in files @file{hoshi.db} etc. which are
automatically build from the files @file{hoshi.sgf} etc. These comprise
our small Joseki library.
@item patterns in @file{owl_attackpats.db}, @file{owl_defendpats.db}
and @file{owl_vital_apats.db}. These generate moves for the owl
code (@pxref{The Owl Code}).
@item Connection patterns in @file{conn.db} (@pxref{Connections Database})
@item Influence patterns in @file{influence.db} and @file{barriers.db}
@item eye patterns in @file{eyes.db} (@pxref{Eyes}).
The following list contains, in addition to distributed source files
some intermediate automatically generated files such as @file{patterns.c}.
These are C source files produced by "compiling" various pattern
databases, or in some cases (such as @file{hoshi.db}) themselves
automatically generated pattern databases produced by "compiling"
joseki files in Smart Game Format.
Database of connection patterns.
Automatically generated file, containing connection
patterns in form of struct arrays, compiled by @command{mkpat}
Automatically generated file, containing eyeshape
patterns in form of struct arrays, compiled by @command{mkpat}
Header file for @file{eyes.c}.
Database of eyeshape patterns. @xref{Eyes}, for
These are helper functions to assist in evaluating
Smart Game Format file containing 4-4 point openings
Automatically generated database of 4-4 point opening
patterns, make by compiling @file{hoshi.sgf}
Joseki compiler, which takes a joseki file in
Smart Game Format, and produces a pattern database.
Smart Game Format file containing 3-4 point openings
Automatically generated database of 3-4 point opening
patterns, make by compiling @file{komoku.sgf}
Pattern compiler for the eyeshape databases. This
program takes @file{eyes.db} as input and produces @file{eyes.c}
Pattern compiler for the move generation and connection
databases. Takes the file @file{patterns.db} together with
the autogenerated Joseki pattern files @file{hoshi.db}, @file{komoku.db},
@file{sansan.db}, @file{mokuhadzushi.db}, @file{takamoku.db} and produces
@file{patterns.c}, or takes @file{conn.db} and produces @file{conn.c}.
@item @file{mokuhazushi.sgf}
Smart Game Format file containing 5-3 point openings
@item @file{mokuhazushi.db}
Pattern database compiled from mokuhadzushi.sgf
Smart Game Format file containing 3-3 point openings
Pattern database compiled from @file{sansan.sgf}
@item @file{takamoku.sgf}
Smart Game Format file containing 5-4 point openings
Pattern database compiled from takamoku.sgf.
Pattern data, compiled from patterns.db by mkpat.
Header file relating to the pattern databases.
@item @file{patterns.db} and @file{patterns2.db}
These contain pattern databases in human readable form.
@section Coding styles and conventions
@subsection Coding Conventions
Please follow the coding conventions at:
@url{http://www.gnu.org/prep/standards_toc.html}
Please preface every function with a brief description
Please help to keep this Texinfo documentation up-to-date.
A function @code{gprintf()} is provided. It is a cut-down
@code{printf}, supporting only @code{%c}, @code{%d},
@code{%s}, and without field widths, etc. It does, however,
add some useful facilities:
Takes two parameters, and displays a formatted board co-ordinate.
Trace messages are automatically indented to reflect
the current stack depth, so it is clear during read-ahead
when it puts a move down or takes one back.
@quotation As a workaround, @code{%o} at the beginning of the
format string suppresses the indentation.
Normally @code{gprintf()} is wrapped in one of the following:
Print the message if the 'verbose' variable > 0.
(verbose is set by @command{-t} on the command line)
@code{DEBUG(flags, fmt, ...)}:
While @code{TRACE} is intended to afford an overview
of what GNU Go is considering, @code{DEBUG} allows occasional
in depth study of a module, usually needed when something
goes wrong. @code{flags} is one of the @code{DEBUG_*} symbols in
@file{engine/gnugo.h}. The @code{DEBUG} macro tests to
see if that bit is set in the @code{debug} variable, and prints
the message if it is. The debug variable is set using the
@command{-d} command-line option.
The variable @code{verbose} controls the tracing. It
can equal 0 (no trace), 1, 2, 3 or 4 for increasing
levels of tracing. You can set the trace level at
the command line by @option{-t} for @code{verbose=1},
@option{-t -t} for @code{verbose=2}, etc. But in
practice if you want more verbose tracing than level
1 it is better to use GDB to reach the point where
you want the tracing; you will often find that the
variable @code{verbose} has been temporarily set to zero
and you can use the GDB command @command{set var verbose=1}
to turn the tracing back on.
Related to tracing are assertions. Developers are strongly encouraged
to pepper their code with assertions to ensure that data structures
are as they expect. For example, the helper functions make assertions
about the contents of the board in the vicinity of the move they
@code{ASSERT()} is a wrapper around the standard C @code{assert()}
function. In addition to the test, it takes an extra pair of parameters
which are the co-ordinates of a "relevant" board position. If an
assertion fails, the board position is included in the trace output, and
@code{showboard()} and @code{popgo()} are called to unwind and display
We have adopted the convention of putting the word FIXME
in comments to denote known bugs, etc.
@node Navigating the Source
@section Navigating the Source
If you are using Emacs, you may find it fast and convenient to use
Emacs' built-in facility for navigating the source. Switch to the
root directory @file{gnugo-3.6/} and execute the command:
find . -print|grep "\.[ch]$" | xargs etags
This will build a file called @file{gnugo-3.6/TAGS}. Now to
find any GNU Go function, type @command{M-.} and enter the
command which you wish to find, or just @command{RET} if
the cursor is at the name of the function sought.
The first time you do this you will be prompted for the location
of the TAGS table. Enter the path to @file{gnugo-3.6/TAGS}, and
henceforth you will be able to find any function with a minimum