Contents

• Go Programs Overview
• First Go programs - Go0 and Go1
• Go2 (GoRandom), optimized version of Go1
• GoHeuristicPlayer, contains a Go player using a simple "average liberties" heuristic
• Go3 (GoFlatMC), simulation-based Go player
• Go4 (GoFlatMCProb), Simulation-based Go player with probabilistic playout policy Go4
• Go5 (GoMCTS), MCTS-based Go player
• GoAlpha, Alpha Zero training and Go player

Go Programs Overview

Go programs are written in Python, and communicate via GTP. You can run them from the command line, from an automated tool, or from a user interface such as GoGui.

First Go programs - Go0 and Go1

Download go0and1.tgz and unpack it. The go0and1 directory contains the following files:

• Go0.py - the Go0 class and the main function of program Go0
• Go1.py - the Go1 class and the main function of program Go1
• board_base.py - Constants and other basic definitions used by board.py and throughout the code
• board.py - the class GoBoard implementing a basic Go board
• board_util.py - board-related utility functions
• engine.py - common code for Go0, Go1 and later Go engines
• gtp_connection.py - the class GtpConnection implementing a Go Text Protocol (GTP) connection
• gtp_connection_go1.py - example for how to extend a GTP connection with more commands
• play_games.py - play 100 self-play games on 7x7 for profiling
• profile_Go1.py - used for profiling - see lecture 6
• Several shell scripts *.sh for testing
• A number of unit test files, test_*, for example test_board.py - incomplete but functional unit tests for board.py
• A number of GTP test files, *.tst, used for tests

Go2, Optimized almost-random Go Player

• Download go2.tgz and unpack it. The go2 directory contains the program Go2, which is the optimized version of Go1 as discussed in Lecture 6. The main changes are in board.py and many other files are unchanged from go0and1. You can run profile_Go2.py and compare it with the Go1 profile.

GoHeuristicPlayer, contains a Go player using a simple "average liberties" heuristic

• Download go_heuristic_player.tgz and unpack it.
• The go_heuristic_player directory currently contains one heuristics-based player, average_player.py.

Go3, Simulation-based Go player

• Download go3.tgz. The Go3 program is a "flat Monte Carlo" simulation-based Go player which supports several different options and display commands. It does only simulations, no tree search.
• Go3 implements several variations of simulation-based Go players. All choose their best move based on its success in simulations.
• Go3 implements two different simulation policies.
• Go3 also implements two different move selection algorithms at the root.
• See the Go3 documentation for command line options, GTP commands etc.

Go4, Simulation-based Go player with probabilistic playout policy

• Download go4.tgz
• This player extends Go3 by providing a probabilistic playout policy. It is slower than Go3 but performs more informed simulations.

Go5, MCTS-based Go player

• Download go5.tgz
• This player replaces the "flat MC" algorithm of Go3 and Go4 by Monte Carlo Tree Search. Given enough time, the search leads to improved play.
• Note that python is very slow, and there is some overhead related to the search tree, so depending on details of a game and its implementation, Go3 or Go4 could actually be stronger than Go5 for low time limits.

GoAlpha, Alpha Zero training and Go player

• Download go_alpha.tgz
• This code implements Alpha Zero training, and includes a player with a pre-trained neural net.
• Also see Lecture 24 for a description and instructions.

Last modified: Martin Müller