c-gomoku-cli is a command line interface for gomoku/renju engines that support Gomocup protocol. It is a derived project from c-chess-cli, originally authored by Lucas Braesch (lucasart).
Different from computer chess community where a lot of interface tools have been developed in the passed 20 years, computer gomoku/renju is still lack of tools for engine-vs-engine matches. Most of the existing tools are GUI based, which is not suitable for distributed systems like high performance clusters. The motivation of c-gomoku-cli is to provide an open-source command line tool that can run large scale engine-vs-engine matches on multi-core workstations or cluster systems in parallel. The large-scale engine-vs-engine testing is common approach for quantifying the strength improvements of computer game engines, like chess. With this tool, engine designers can concisely measure the new engine ELO gain by letting it play with the old version for enough number of games.
Instead of sticking on the C language to avoid the dependency hell in c-chess-cli, C++ is chosen to be the primary programming language in this project, because my target is to build a usable tool and C++ contains some features that can bring more convenience to achieving this target. Similar to c-chess-cli, c-gomoku-cli is primarily developed for Unix/Linux, it should work on all POSIX operating systems, including MacOS and Android. Windows support is added by Haobin (dhbloo).
Under the root directory of project, run cd src, then run make.
c-gomoku-cli [-each [eng_options]] -engine [eng_options] -engine [eng_options] ... [options]c-gomoku-cli -each tc=180/30 \
-engine name=Wine cmd=./example-engine/pbrain-wine \
-engine name=Rapfi cmd=./example-engine/pbrain-rapfi1 \
-rule 0 -boardsize 20 -rounds 2 -games 4000 -debug -repeat \
-concurrency 8 -drawafter 200 \
-pgn my_games.pgn -sgf my_games.sgf -msg my_games.txt \
-openings file=opening_examples/offset_freestyle_20x20.txt order=randomengine OPTIONS: Add an engine defined by OPTIONS to the tournament.each OPTIONS: Apply OPTIONS to each engine in the tournament.concurrency N: Set the maximum number of concurrent games to N (default value 1).drawafter N: Adjudicate the game as a draw, if the number of moves in one game reaches N ply. N must be greater then 0 to be effective.rule RULE: Set the game rule with Gomocup rule code RULE.
RULE=0: Play with gomoku rule and winner wins by five or longer connection.RULE=1: Play with gomoku rule but winner only wins by exact-5 connection (longer connections will be ignored).RULE=4: Play with renju rule.boardsize SIZE: Set the board size to SIZE X SIZE. Valid SIZE range is [5..22]. games N: Play N games per encounter (default value 1). This value should be set to an even number in tournaments with more than two players to make sure that each player plays an equal number of games with black and white colors.rounds N: Multiply the number of rounds to play by N (default value 1). This only makes sense to use for tournaments with more than 2 engines.gauntlet: Play a gauntlet tournament (first engine against the others). The default is to play a round-robin (plays all pairs).
n=2 engines, both gauntlet and round-robin just play the number of -games specified.n>2: G(e1, ..., en) = G(e1, e2) + G(e1, e3) + ... + G(e1, en). There are n-1 pairs.n>2: RR(e1, ..., en) = G(e1, ..., en) + RR(e2, ..., en). There are n(n-1)/2 pairs.-rounds repeats the tournament -rounds times. The number of games played for each pair is therefore -games * -rounds.loseonly: In a gauntlet tournament, only save games, messages and samples that first engine loses. This option is only effective when specifying gauntlet.repeat: Repeat each opening twice, with each engine playing both sides. transform: Transform openings by using rotating and flip. There are 8 types of transform (identity, rotate90, rotate180, rotate270, flipX, flipY, flipXY, flipYX). After using all openings each time, a new transform type is used, and this process repeats for all transform types.sprt [elo0=E0] elo1=E1 [alpha=A] [beta=B]: Performs a Sequential Probability Ratio Test for H1: elo=E1 vs H0: elo=E0, where alpha is the type I error probability (false positive), and beta is type II error probability (false negative). Default values are elo0=0, and alpha=beta=0.05. This can only be used in matches between two players.log: Write all I/O communication with engines to file(s). This produces c-gomoku-cli.id.log, where id is the thread id (range 1..concurrency). Note that all communications (including error messages) starting with [id] mean within the context of thread number id, which tells you which log file to inspect (id = 0 is the main thread, which does not product a log file, but simply writes to stdout).debug: Turn on debug mode. In debug mode, more detailed information about game and engines will be printed, and -log will also be turned on automatically.sendbyboard: Send full position using BOARD command before each move. If not specified, continuous position are sent using TURN. Some engines might behave differently when receiving BOARD rather than TURN.fatalerror: Consider "engine crashed before answering to START", "engine timeout after tolerance before answering to START", "engine output ERROR before answering to START", "engine crashed before answering to MOVE", "engine timeout after tolerance before answering to MOVE" as fatal error, which causes c-gomoku-cli to terminate with a failure exit code. By default this is turned off thus such engine failure is considered as crash loss or time loss (Error messages will still be printed to stderr).openings file=FILE [type=TYPE] [order=ORDER] [srand=N]:
FILE, in TYPE format. type can be offset (default value) or pos. See "Openings File Format" section below about details of different formats.order can be random or sequential (default value).srand sets the seed of the random number generator to N. The default value N=0 will set the seed automatically to an unpredictable number. Any non-zero number will generate a unique, reproducible random sequence.pgn FILE: Save a dummy game to FILE, in PGN format. PGN format is for chess games. We replace the moves with some random chess moves but only keep the game result and player names. This dummy PGN file can be input by BayesianElo to compute ELO scores.sgf FILE: Save a game to FILE, in SGF format.msg FILE: Save engine messages to FILE, in TXT format. Messages in each games are grouped by game index.sample. See below.
cmd=COMMAND: Set the command to run the engine.
/ is contained in COMMAND. For example, cmd=../Engines/Wine2.0, will run ./Wine2.0 from ../Engines. If no / is found, the command is executed as is. Without /, for example cmd=demoengine will run demoengine, which only works if demoengine command was in PATH."cmd=../fooEngine -foo=1". Note that the "" are needed here, for the command line interpreter to parse the whole string as a single token.\). If you would like to use backslash in engine path on Windows, use a double backslash (\\) to replace all original backslash in path.name=NAME: Set the engine's name. If omitted, the name will be taken from the ABOUT values sent by the engine.
tc=TIMECONTROL: Set the time control to TIMECONTROL. The format is match_time/turn_time+increment or match_time, where match_time is the total time of this match (in seconds), turn_time is the max time limit per move (in seconds), and increment is time increment per move (in seconds). If turn_time is omitted, then it will be the same with match_time by default. If increment is omitted, then it will be set to 0 by default. If match_time is 0, then there will no limit on match time and turn_time will be the only limitation.
depth=N: Depth limit per move. This is an extension option[^1], may not be supported by all engines.
nodes=N: Node limit per move (N is recommended to be have a granularity more than 1000). This is an extension option[^1], may not be supported by all engines.
maxmemory=MAXMEMORY: Set the max memory to MAXMEMORY bytes. Default memory limit is 350MB (same as Gomocup) if omitted.
thread=N: Number of threads a engine can use. Default value is 1. This is an extension option[^1], may not be supported by all engines.
tolerance=N: Tolerance (in seconds) to determine when an engine hangs (which is an unrecoverable error at this point). Default value is N=3.
option.O=V: Set a raw protocol info. Command INFO [O] [V] will be sent to the engine before each game starts.
[^1]: Yixin-Board extension protocol: https://github.com/accreator/Yixin-protocol/blob/master/protocol.pdf
So far c-gomoku-cli only accept openings in plaintext format (*.txt). In a plaintext opening file, each line is an opening position. Currently there are two notation types for a position: offset and pos.
Offset opening notation is denoted by the move sequence: <black_move>, <white_move>, <black_move>... . Each move is separated with a comma "," and a space "". The last move will not be followed by any comma or space. Each move is in format <x-offset>,<y-offset>, no space in between. Note that here each coordinate is the offset from the center of the board, which is different from the move coordinate of Gomocup protocol where (0,0) is up-left corner of the board. Offset value can be negative.
Assuming that a board is in size SIZE, the conversion between plaintext offsets <x-offset>,<y-offset> and Gomocup move <x>,<y> is:
HALF_SIZE = floor(SIZE / 2)
<x> = <x-offset> + HALF_SIZE
<y> = <y-offset> + HALF_SIZEThe following is an example position in "offset" notation:
8,-3, 6,-4, 5,-4, 4,-3, 2,-8, -1,-5You can see more examples from Gomocup 2020 result page (download the results+openings from the link at the very bottom of that page).
Pos opening notation is denoted by a sequence of move position, with no space or any forms of delimiter in between: <black_move><white_move><block_move>... . Each move is in format <x-coord><y-coord>, while <x-coord> is a letter starting from "a", <y-coord> is a number starting from "1". Coordinate is the offset from the upper-left corner of the board. This notation is only suitable for board size less than 27.
The following is an example position in "pos" notation:
b7d6e6f7h2k5This notation is common among many Gomoku/Renju applications. (For example, you can acquire a pos notation text by using "getpos" command in Yixin-Board).
Sampling is used to record various position and engine outputs in a game, as well as the final game result. These can be used as training data, which can be used to fit the parameters of a gomoku engine evaluation, otherwise known as supervised learning. Sample record is usually in binary format easy for engine to process, meanwhile a human readable and easily parsable CSV file can also be generated. Note that only game which result is not "win by time forfeit" or "win by opponent illegal move" will be recorded.
Syntax is -sample [freq=%f] [format=csv|bin|bin_lz4] [file=%s]. Example -sample freq=0.25 format=csv file=out.csv.
freq is the sampling frequency (floating point number between 0 and 1). Defaults to 1 if omitted.file is the name of the file where samples are written. Defaults to sample.[csv|bin|bin.lz4] if omitted.format is the format in which the file is written. Defaults to csv, which is human readable: Position,Move,Result. Position is the board position in "pos" notation. Move is the move in "pos" notation output by the engine. Result is the game outcome from perspective of current side to move, values for Result are 0=loss, 1=draw, 2=win. For binary format bin see the section below for details. bin_lz4 is the same as bin format, but the whole file stream is compressed using LZ4 to save disk space (This is suitable for huge training dataset containing millions of positions). Engines are recommended to use LZ4 "Auto Framing" API (example) to decompress the training data.Binary format uses variable length encoding show below, which is easy to parse for engines. Each entry has a length of 4+ply bytes. Position is represented by a move sequence that black plays first. Move sequence is guaranteed to have the same order as the actual game record.
struct Entry {
uint16_t result : 2; // game outcome: 0=loss, 1=draw, 2=win (side to move pov)
uint16_t ply : 9; // current number of stones on board
uint16_t boardsize : 5; // board size in [5-22]
uint16_t rule : 3; // game rule: 0=freestyle, 1=standard, 4=renju
uint16_t move : 13; // move output by the engine
uint16_t position[ply]; // move sequence that representing a position
};Each move is represented by a 16bit unsigned integer. It's lower 10 bits are constructed with two index x and y using uint16_t move = (x << 5) | y;. Below is a code snippet that does transform between the packed move and two coordinates.
uint16_t Move(int x, int y) { return (x << 5) | y; }
int CoordX(uint16_t move) { return (move >> 5) & 0x1f; }
int CoordY(uint16_t move) { return move & 0x1f; }Thanks to lucasart for developing the c-chess-cli project. His prior work provides a perfect starting point for the development of c-gomoku-cli. Thanks to Haobin for contributing the support on Windows. It makes c-gomoku-cli avaliable to all Windows based engines.
External library used:
Lucas Braesch (lucasart) Chao Ma (nkg114mc) Haobin Duan (dhbloo)