Human vs. Computer Go: Review and Prospect

The Google DeepMind challenge match in March 2016 was a historic achievement for computer Go development. This article discusses the development of computational intelligence (CI) and its relative strength in comparison with human intelligence for the game of Go. We first summarize the milestones achieved for computer Go from 1998 to 2016. Then, the computer Go programs that have participated in previous IEEE CIS competitions as well as methods and techniques used in AlphaGo are briefly introduced. Commentaries from three high-level professional Go players on the five AlphaGo versus Lee Sedol games are also included. We conclude that AlphaGo beating Lee Sedol is a huge achievement in artificial intelligence (AI) based largely on CI methods. In the future, powerful computer Go programs such as AlphaGo are expected to be instrumental in promoting Go education and AI real-world applications. I. Computer Go Competitions The IEEE Computational Intelligence Society (CIS) has funded human vs. computer Go competitions in IEEE CIS flagship conferences since 2009. Fig. 1 shows the year and the location of the conferences. The descriptions of competitions held from 1998 to 2016 are listed in detail in an online version of this article [1-8]. The handicaps for the human vs. computer 19×19 game have been decreased from 29 in 1998 to 0 in 2016. The skill of amateur players in Go is ranked according to kyu (K) in the lower tier, where a smaller number stands for stronger playing skill (with 1K being the highest skill level), and dan (D) in the higher tier, where a larger number stands for stronger playing skill. Professional Go players are ranked entirely in dan, abbreviated with the letter P (e.g. Lee Sedol is ranked at 9P). In the amateur level, each difference in rank roughly translates to a single stone of handicap (H), where the weaker player is allowed to place an additional stone on the board prior to play to even out the game. The skill difference between professional ranks is much less than one stone for every rank difference. Go is typically played on 19×19 size boards, but 9×9 size boards are also common for beginners. The complexity of the 9×9 game is far less than the standard game, and the 9×9 game had been one of the interim goals for computer Go programs. Go is a game that is inherently biased for the first player to play, Black. To compensate for this first player advantage, White is awarded additional points at the end of the game, which is referred to as komi. The related statistics for the IEEE CIS human vs. computer Go competitions are listed in the online version of this article [8]. It is worth noting that with a komi of 7.5, White may end up with an advantage in both 9×9 games and handicapped 19×19 games, regardless of whether White is played by humans or computers. Fig. 2 shows the certificate awarded to the MoGoTW program (16cores / 48GB / 9×9) by the Taiwan Go Association in 2010 for playing at a 3D amateur level * Corresponding author: Chang-Shing Lee (E-mail: [email protected]). This article is accepted and will be published in IEEE Computational Intelligence Magazine in August 2016 2 [1]. In this article, we attempt to demonstrate the massive barrier that competing programs need to overcome to achieve comparable performance to AlphaGo. For more information, eleven programs (Aya, CGI, Coldmilk/Jimmy, Erica, Fuego, MoGo/MoGoTW, Many Faces of Go, Pachi, and Zen) from 7 countries that have participated in past IEEE CIS conferences are listed in alphabetical order in the online version of this article [8]. Fig. 1. Past human vs. computer Go competitions in IEEE CIS flagship conferences. Fig. 2. Amateur 3D certificate awarded to MoGoTW in 2010.