Pentobi GTP Interface
Author: Markus Enzenberger
This document describes the text-based interface to the engine
of the Blokus program
Pentobi.
The interface is an adaption of the
Go Text Protocol
(GTP) and allows controller programs to use the engine in an automated
way without the GUI.
The most recent version of this document can be found in the source
code distribution of Pentobi in the folder pentobi/doc/gtp.
Go Text Protocol
The Go Text Protocol is a simple text-based protocol.
The engine reads single-line commands from its standard input stream and
writes multi-line responses to its standard output stream.
The first character of a response is a status character:
= for success, ? for failure, followed by the actual
response.
The response ends with two consecutive newline characters.
See the
GTP specification for details.
Controllers
To use the engine from a controller program, the controller typically
creates a child process by running pentobi-gtp and then sends
commands and receives responses through the input/output streams of the
child process.
How this is done, depends on the platform (programming language and/or
operating system).
In Java, for example, a child process can be created with
java.lang.Runtime.exec().
Note that the input/output streams of child processes are often fully
buffered.
You should explicitly flush the output stream after sending a command.
Another caveat is that pentobi-gtp writes debugging information
to its standard error stream.
On some platforms the standard error stream of the child process
is automatically connected to the standard error stream of its parent
process.
If not (this happens for example in Java), the controller needs
to read everything from the standard error stream of the child process.
This can be done for example by running a separate thread in the parent
process that has a simple read loop, which writes everything that it
reads to its own standard error stream or discards it.
Otherwise the child process will block as soon as the buffer for its
standard error stream is full.
Alternatively, you can disable debugging output of pentobi-gtp
with the command line option --quiet, but it
is generally better to assume that a GTP engine writes text to standard
error.
An example for a controller written in Python is
tools/twogtp/twogtp.py, which starts two GTP engines and
plays a number of Blokus games between them.
Building
Since the GTP engine is a developer tool, building it is not enabled
by default.
To enable it, run cmake with the option
-DPENTOBI_BUILD_GTP=ON.
After building, there will be an executable in the build directory named
src/pentobi_gtp/pentobi-gtp.
The GTP engine requires only standard C++ and has no dependency on other
libraries like Qt, which is needed for the GUI version of Pentobi.
If you only want to build the GTP engine, you can disable building the
GUI with -DPENTOBI_BUILD_GUI=OFF.
Options
The following command-line options are supported by pentobi-gtp:
- --book file
-
Specify a file name for the opening book.
Opening books are blksgf files containing trees, in which moves that
Pentobi should select are marked as good moves with the
corresponding SGF property (see the files in src/books).
If no opening book is specified and opening books are not disabled,
pentobi-gtp will automatically search for an opening book
for the current game variant in the directory of the executable
using the same file name conventions as in src/books.
If no such file is found it will print an error message to
standard error and disable the use of opening books.
- --config,-c file
-
Load a file with GTP commands and execute them before starting the
main loop, which reads commands from standard input.
This can be used for configuration files that contain GTP commands
for setting parameters of the engine (see below).
- --color
-
Use ANSI escape sequences to colorize the text output of boards
(for example in the response to the showboard command or
with the --showboard command line option).
- --cputime
-
Use CPU time instead of wall time for time measurement.
Currently, there is no way to make Pentobi play with time limits,
the levels are defined by the number of simulations in the MCTS
search, so this affects only the debugging output, which prints
the time used after each search.
- --game,-g variant
-
Specify the game variant used at start-up.
Valid arguments are classic, classic_2, duo, trigon, trigon_2,
trigon_3, junior or the abbreviations c, c2, d, t, t2, t3, j.
By default, the initial game variant is classic.
The game variant can also be changed at run-time with a GTP command.
If only a single game variant is used, it is slightly faster and
saves memory if the engine is started in the right variant compared
to having it start with classic and then changing it.
- --help,-h
-
Print a list of the command-line options and exit.
- --level,-l n
-
Set the level of playing strength to n.
Valid values are 1 to 8.
Higher levels are allowed, but currently untested and unsupported
(they will increase the search time but may only make sense if the
amount of memory is increased with the option --memory and
compilation is done with the cmake option
-DLIBPENTOBI_MCTS_FLOAT_TYPE=double).
- --memory n
-
Allocate n bytes of memory for the search trees.
If not specified, Pentobi will currently allocate a third of the
physical memory available on the system but not more than 1 GB.
Increasing the memory may slightly increase the playing strength
at higher levels.
If only lower levels are used, the amount of memory can be reduced.
If Pentobi runs out of memory (typically in early game positions
in game variant Duo), it will prune the search tree and print
debugging messages about the pruning to standard error.
The memory can be reduced if pruning does not happen more than a
few times during a move generation.
- --seed,-r n
-
Use n as the seed for the random generator.
Specifying a random seed will make the move generation deterministic
as long as the search is single-threaded.
- --showboard
-
Automatically write a text representation of the current position
to standard error after each command that alters the position.
- --nobook
-
Disable the use of opening books.
- --noresign
-
Disable resignation.
If resignation is desabled, the genmove command will
never respond with resign.
Resignation can speed up the playing of test games if only the
win/loss information is wanted.
- --quiet,-q
-
Do not print any debugging messages, errors or warnings to standard
error.
- --threads n
-
Use n threads during the search.
Note that the default is 1, unlike in the GUI version of Pentobi,
which sets the default according to the number of hardware threads
(CPUs, cores or virtual cores) available on the current system.
The reason is that, for example, using 2 threads makes the search
twice as fast but may lose a bit of playing strength compared
to the single-threaded search.
Therefore, if the GTP engine is used to play many test games with
twogtp.py (which supports playing games in parallel), it is better
to play the games with single-threaded search in parallel than with
multi-threaded search sequentially.
Using a large number of threads (e.g. more than 4 on game variants
with a small board or more than 8 on large boards) is untested
and might reduce the playing strength significantly compared
to the single-threaded search.
- --version,-v
-
Print the version of Pentobi and exit.
Commands
Standard Commands
The following GTP commands have the same or an equivalent meaning
as specified by the GTP standard.
Colors or players in arguments or responses are represented as in the
property IDs of blksgf files (B, W if two players
or colors; 1, 2, 3, 4 if more than
two).
Moves in arguments or responses are represented as in the move property
values of blksgf files.
See the specification for
Pentobi SGF files
for details.
- all_legal color
-
List all legal moves for a color.
- clear_board
-
Clear the board and start a new game in the current game variant.
- final_score
-
Get the score of a final board position.
In two-player game variants, the format of the response is as in the
result property in the SGF standard for the game of Go
(e.g. B+2 if the first player wins with two points,
or 0 for a draw).
In game variants with more than two players, the response is a
list of the points for each player
(e.g. 64 69 70 40).
If the current position is not a final position, the response is
undefined.
- genmove color
-
Generate and play a move for a given color in the current position.
If the color has no more moves, the response is pass.
If resignation is not disabled, the response is resign
if the players is very likely to lose.
Otherwise the response is the move.
- known_command command
-
The response is true if command is a GTP command
supported by the engine, false otherwise.
- list_commands
-
List all supported GTP commands, one command per line.
- loadsgf file [move_number]
-
Load a board position from a blksgf file with name file.
If move_number is specified, the board position will be
set to the position in the main variation of the file before
the move with the given number was played, otherwise to the last
position in the main variation.
- name
-
Return the name of the GTP engine (Pentobi).
- play color move
-
Play a move for a given color in the current board position.
- protocol_version
-
Return the version of the GTP protocol used (currently 2).
- quit
-
Exit the command loop and quit the engine.
- reg_genmove color
-
Like the genmove command, but only generates a move and does
not play it on the board.
- showboard
-
Return a text representation of the current board position.
- undo
-
Undo the last move played.
- version
-
Return the version of Pentobi.
Generally Useful Extension Commands
- cputime
-
Return the CPU time used by the engine since the start of the program.
- cputime_diff
-
Return the CPU time used by the engine since the last call
of cputime_diff or start of the program if cputime_diff has nto been
called yet.
- g
-
Shortcut for the genmove command with the color argument
set to the current color to play.
- get_place color
-
Get the place of a given color in the list of scores in a final
position (e.g. in game variant Classic, 1 is the place with the
highest score, 4 the one with the lowest, if all players have
a different score).
If some colors have the same score, they share the same place
and the string shared is appended to the place number.
- get_value
-
Get an estimated value of the board position from the view point
of the color of the last generated move.
The return value is a win/loss estimation between 0 (loss) and
1 (win) as produced by the last search performed by the engine.
This command should only be used immediately after a
reg_genmove or genmove command, otherwise the
result is undefined.
The value is not very meaningful at the lowest playing levels.
Note that no searches are performed if the opening book is used
for a move generation and there is currently no way to check if this
was so.
Therefore, the opening book should be disabled if the
get_value command is used.
- p move
-
Shortcut for the play command with the color argument set
to the current color to play.
- param [key value]
-
Set or query parameters specific to the Pentobi engine that can be
changed at run-time.
If no arguments are given, the response is a list of the
current value with one key/value pair per line, otherwise
the parameter with the given key will be set to the given value.
Generally useful parameters are:
- avoid_symmetric_draw 0|1
-
In some game variants (Duo, Trigon_2), the second player can
enforce a tie by answering each move by its symmetric counterpart
if the first players misses the opportunity to break the symmetry
in the center.
Technically, exploiting this mistake by the first player is a
good strategy for the second player because a draw is a good
result considering the first-play advantage.
However, playing symmetrically could be considered bad style,
so this behavior is avoided (value 1) by default.
- fixed_simulations n
-
Use exactly n MCTS simulations during a search.
By default, the search engine uses levels, which determine
how many MCTS simulations are run during a search, but as a
function that increases with the move number (because the
simulations become much faster at the end of the game).
For some experiments, it can be desirable to use a fixed number
of simulations for each move.
If this number is specified, the playing level is ignored.
- level n
-
Set the playing level to n.
See the command line option --level for possible values.
- use_book 0|1
-
Enable or disable the opening book.
The other parameters are only interesting for developers.
- param_base [key value]
-
Set or query basic parameters that are not specific to the Pentobi
engine.
If no arguments are given, the response is a list of the
current value with one key/value pair per line, otherwise
the parameter with the given key will be set to the given value.
- accept_illegal 0|1
-
Accept move arguments to the play command that violate
the rules of the game.
If disabled, the play command will respond with an error,
otherwise it will perform the moves.
- resign 0|1
-
Allow the engine to respond with resign to the
genmove command.
- set_game variant
-
Set the current game variant and clear the board.
The argument is the name of the game variant as in the game property
value of blksgf files (e.g. Blokus Duo,
see the specification for
Pentobi SGF files
for details).
- set_random_seed n
-
Set the seed of the random generator to n.
See the documentation for the command-line option --seed.
Extension Commands for Developers
The remaining commands are only interesting for developers.
See Pentobi's source code for details.
Example
The following GTP session queries the engine name and version,
plays and generates a move in game variant Duo and shows the resulting
board position.
Commands are printed in bold, responses in normal text.
$ ./pentobi-gtp --quiet
name
= Pentobi
version
= 7.1
set_game Blokus Duo
=
play b e8,d9,e9,f9,e10
=
genmove w
= i4,h5,i5,j5,i6
showboard
=
A B C D E F G H I J K L M N
14 . . . . . . . . . . . . . . 14 *Blue(X): 5
13 . . . . . . . . . . . . . . 13 1 F L5 N P T5 U V5 W Y
12 . . . . . . . . . . . . . . 12 Z5 I5 O T4 Z4 L4 I4 V3 I3 2
11 . . . . . . . . . . . . . . 11
10 . . . . X . . . . . . . . . 10 Green(O): 5
9 . . . X X X . . . . . . . . 9 1 F L5 N P T5 U V5 W Y
8 . . . . X . . . . . . . . . 8 Z5 I5 O T4 Z4 L4 I4 V3 I3 2
7 . . . . . . . . . . . . . . 7
6 . . . . . . . .>O . . . . . 6
5 . . . . . . . O O O . . . . 5
4 . . . . . . . . O . . . . . 4
3 . . . . . . . . . . . . . . 3
2 . . . . . . . . . . . . . . 2
1 . . . . . . . . . . . . . . 1
A B C D E F G H I J K L M N
quit
=