Introduction to Go Programs Developed at CGI Lab

This page introduces the history of Go programs developed in Computer Games and Intelligence (CGI) lab, led by Professor I-Chen Wu, at Department of Computer Science, National Chiao Tung University, Taiwan; and also acknowledges the contributors who are/were the members of the CGI lab, unless specified explicitly. For the current strength of CGI, see KGS account "CGI".

HappyGo (2008-2011)

One year before 2008, Yung-Ler Wang (王永樂) wrote a simple version of Go program, named HappyGo, since his nickname was Happy. Although the program was completely rewritten after 2008 by the following authors, we still used the name HappyGo. Note that we focused on 9x9 Go in this version.

Authors

Chun-Yi Chen (陳俊嶧) 2008-2010 Wrote most of the code of HappyGo, including Monte-Carlo Tree Search (MCTS) and many heuristics.
Tsun-Tao Tso (左存道) 2011-2012 Wrote a solver inside HappyGo, including the design of relevance zones.
I-Chen Wu (吳毅成) 2008-2012 Supervised the whole project.

Features

Competitions

Year Competitions Place
2013 TCGA Go 9x9 2nd place (Silver)
2011 TAAI Go 9x9 5th place
2010 TAAI Go 9x9 2nd place (Silver)
2010 ICGA Go 9x9 8th place
2009 TAAI Go 9x9 4th place

Publications

  1. Chun-Yi Chen, "A Design of Monte-Carlo Computer Go Program", M.S. thesis, National Chiao Tung University, Hsinchu, Taiwan, 2010.
  2. Tsun-Tao Tso, "A Study of Solving Life and Death Problem for 7x7 Kill-All Go", M.S. thesis, National Chiao Tung University, Hsinchu, Taiwan, 2012.

Amigo (2012-2014)

Ting-Fu Liao rewrote the whole Go program that contains a general framework for parallel Monte Carlo Tree Search, which can be used for other games, like Chinese dark chess, NoGo. For generality, the initial version of Amigo by Ting-Fu Liao contained not much Go-specific code, except that it followed the rule-based playouts of HappyGo partially. Note that we still focused on 9x9 Go in this version, but it works for 19x19 due to generality (but very weak). So, we can say that 19x19 had not been developed in this stage.

Authors

Ting-Fu Liao (廖挺富) 2011-2013 Wrote general framework for parallel Monte Carlo Tree Search, it can be use in any game.
Ting-Chu Ho (何庭築) 2012-2014 Assisted on Go-specific knowledge like connection.
I-Chen Wu (吳毅成) 2011-2014 Supervised the whole project.

Features

Competitions

Year Competitions Place
2014 TAAI Go 9x9 2nd place (Silver)
2014 TCGA Go 9x9 2nd place (Silver)
2013 ICGA Go 9x9 3rd place (Bronze)
2013 TCGA Go 9x9 3rd place (Bronze)

Publications

  1. Ting-Fu Liao, "Software Framework for Parallel Monte Carlo Tree Search", M.S. thesis, National Chiao Tung University, Hsinchu, Taiwan, 2013.
  2. Ting-Chu Ho, "The Study of Connections for Computer Go", M.S. thesis, National Chiao Tung University, Hsinchu, Taiwan, 2014.
  3. Ting-Fu Liao, I-Chen Wu, Guan-Wun Chen, Chung-Chin Shih, Po-Ya Kang, Bing-Tsung Chiang, Ting-Chu Ho and Ti-Rong Wu, "A Study of Software Framework for Parallel Monte Carlo Tree Search," the 19th Game Programming Workshop (GPW-2014), Hakone Seminar House, Kanagawa, Japan, November 7-9, 2014.

CGI 1.0 (2015.1-2015.12)

Since Amigo had name conflict with an old Go program, it was renamed to CGI Go Intelligence (abbr. CGI, a kind of recursive acronym). The program was not changed until Ti-Rong Wu (the key developer) as well as some authors (see below) made significant changes in the Go side on top of the framework in 2015. So, CGI is usually referred to the changed version starting from 2015. Since some machine learning methods like M&M were incorporated, the program started working reasonably for 19x19 and the strength grows to 2d or so.

Authors

Ti-Rong Wu (吳廸融) 2015-current Made most significant changes and improvements over Amigo (nearly rewrote the code in the Go side) , mainly including the tuning of pattern features by machine learning method and many detailed work. (The key developer of this CGI version)
Ting-Fu Liao (廖挺富) 2011-2013 Wrote general framework for parallel Monte Carlo Tree Search, which this version is still based on.
Guan-Wen Chen (陳冠文) 2015-current Worked on distributed computing, handling server and worker synchronization.
Chung-Chin Shih (施仲晉) 2015-current Assisted on the implementation of some patterns. To implement knowledges like life and death in Go, also used in KillAll Go.
Li-Cheng Lan (藍立呈) 2015-current Assisted on MCTS including dynamic komi, argument tuning.
Kun-Hao Yeh (葉騉豪) 2015-current To build Go 9x9 opening book by Job-Level System.
I-Chen Wu (吳毅成) 2015-current Supervised the whole project.

Features

Competitions

Human Competitions

Year Competitions (19x19) Human (White) Rule Result
2015 IEEE CIG 2015 Go Chun-Hsun Chou 9p
(周俊勳 職業九段)		 H6, komi: 0.5, 45 minutes each Lose (W+5.5)
2015 IEEE CIG 2015 Go Kai-Hsin Chang 5p
(張凱馨 職業五段)		 H6, komi: 0.5, 45 minutes each Lose (W+Res)
2015 IEEE CIG 2015 Go Li-Chun Yu 1p
(俞俐均 職業一段)		 H6, komi: 0.5, 45 minutes each Win (B+Res)

Computer Competitions

Year Competitions Place
2015 ICGA Go 19x19 4th place
2015 ICGA Go 13x13 4th place
2015 ICGA Go 9x9 3rd place (Bronze)
2015 TCGA Go 19x19 1st place (Gold)
2015 TCGA Go 13x13 1st place (Gold)
2015 TCGA Go 9x9 1st place (Gold)
(Note: TCGA refers to Taiwan Computer Game Association (台灣電腦對局學會), which hosts competitions every year.)

CGI 2.0 (2015.12-2016.8)

As of Dec 2015, we started incorporating Deep Convolutional Neural Network (DCNN) or so-called Deep Learning into CGI, which is called CGI 2.0 or Deep CGI for convenience. Here, we would like to thank and acknowledge Detlef Schmicker for sharing his DCNN data set (with 54% prediction rate), which was used as our initial DCNN data set, though we changed to our own training data in mid-Jan 2016. The following list simply shows the persons involved in this project and is subject to change (Note: Ti-Rong Wu, Li-Cheng Lan, and Guan-Wen Chen are three key developers in the current project).

Authors

Ti-Rong Wu (吳廸融) 2015-current Made most significant changes and improvements over CGI 1.0, and also worked on DCNN.
Li-Cheng Lan (藍立呈) 2015-current Worked on the DCNN architecture, mainly for multi-GPU versions.
Guan-Wen Chen (陳冠文) 2015-current Worked on DCNN training.
Hung-Chun Wu (吳宏君) 2016-current Assisted on DCNN.
Jin-Bo Huang (黃勁博) 2016-current Assisted on DCNN.
Ting-Fu Liao (廖挺富) 2011-2013 the code he wrote for general parallel MCTS framework is still used in this version.
I-Chen Wu (吳毅成) 2015-current Supervised the whole project.

Features

Competitions

Computer Competitions

2016 UEC Cup (19x19 Go) Rank
Day 1 (Preliminary League) All wins for 7 games (including Zen, Ray etc.) 1st place
Day 2 (Final Tournament) Won 2 and lost 2
(Also won the best student award.)		 6th place

Human Competitions

Year Competitions Human (White) Rule Result
2016 IEEE WCCI 2016 Go Chun-Hsun Chou 9p
(周俊勳 職業九段)		 H2, komi: 6.5, 45 minutes each Lose

CGI 3.0 (2016.8-)

As of Aug. 2016, we start the CGI 3.0 project, mainly developing value network (VN), and some other DCNNs, like RL policy network. As of Dec 2015, we started incorporating Deep Convolutional Neural Network (DCNN) or so-called Deep Learning into CGI, which is called CGI 2.0 or Deep CGI for convenience. This version is still being developed. Here, we would like to thank and acknowledge Detlef Schmicker for sharing his DCNN data set (with 54% prediction rate), which was used as our initial DCNN data set, though we changed to our own training data in mid-Jan 2016. The strength of this version grows significantly, but it is still unknown and not final yet, since the project is still ongoing now. The following list simply shows the persons involved in the Deep CGI project and is subject to change (Note: Ti-Rong Wu, Li-Cheng Lan, and Guan-Wen Chen are three key developers in the current project).

Authors

Ti-Rong Wu (吳廸融) 2015-current Made most significant changes and improvements over CGI 1.0, and also worked on DCNN.
Guan-Wen Chen (陳冠文) 2015-current Worked on distributed computing.
Hung-Chun Wu (吳宏君) 2016-current Worked on DCNN training.
Dong-Yi Lai (賴東億) 2016-current Worked on DCNN training.
Peter Wu (吳慈仁) Current To help customization.
Some more Current To help the development.
I-Chen Wu (吳毅成) 2015-current Supervised the whole project.

Features

Competitions

Computer Competitions

Human Competitions

Publications

  1. Ti-Rong Wu, I-Chen Wu, Guan-Wun Chen, Ting-han Wei, Tung-Yi Lai, Hung-Chun Wu, Li-Cheng Lan, "Multi-Labelled Value Networks for Computer Go", arXiv:1705.10701, 2017.

More