Glossary

GTP

The Go Text Protocol

A communication protocol used to control Go-playing programs. Gomill uses only GTP version 2, which is specified at http://www.lysator.liu.se/~gunnar/gtp/gtp2-spec-draft2/gtp2-spec.html.

(As of February 2011, the specification describes itself as a draft, but it has remained stable for several years and is widely implemented.)

SGF

The Smart Game Format

A text-based file format used for storing Go game records.

Gomill uses version FF[4], which is specified at http://www.red-bean.com/sgf/index.html.

jigo

A tied game (after komi is taken into account).

komi

Additional points awarded to White in final scoring.

simple ko

A Go rule prohibiting repetition of the immediately-preceding position.

superko

A Go rule prohibiting repetition of preceding positions.

There are several possible variants of the superko rule. Gomill does not enforce any of them.

controller

A program implementing the ‘referee’ side of the GTP protocol.

The GTP protocol can be seen as a client-server protocol, with the controller as the client.

engine

A program implementing the ‘playing’ side of the GTP protocol.

The GTP protocol can be seen as a client-server protocol, with the engine as the server.

player

A GTP engine, together with a particular configuration.

competition

An ‘event’ consisting of multiple games managed by the Gomill ringmaster (either a tournament or a tuning event).

tournament

A competition in which the ringmaster plays games between predefined players, to compare their strengths.

playoff

A tournament comprising many games played between fixed pairings of players.

all-play-all

A tournament in which games are played between all pairings from a list of players.

matchup

A pairing of players in a tournament, together with its settings (board size, komi, handicap, and so on)

tuning event

A competition in which the ringmaster runs an algorithm which adjusts player parameters to try to find the values which give strongest play.

Bandit problem

A problem in which an agent has to repeatedly choose between actions whose value is initially unknown, trading off time spent on the action with the best estimated value against time spent evaluating other actions.

See http://en.wikipedia.org/wiki/Multi-armed_bandit

UCB

Upper Confidence Bound algorithms

A family of algorithms for addressing bandit problems.

UCT

Upper Confidence bounds applied to Trees.

A variant of UCB for bandit problems in which the actions are arranged in the form of a tree.

See http://senseis.xmp.net/?UCT.