Monte-Carlo Tree Search

representation of the game. It was programmed in LISP. Further use of abstraction was also studied by Friedenbach (1980). The combination of search, heuristics, and expert systems led to the best programs in the eighties. At the end of the eighties a new type of Go programs emerged. These programs made an intensive use of pattern recognition. This approach was discussed in detail by Boon (1990). In the following years, different AI techniques, such as Reinforcement Learning (Schraudolph, Dayan, and Sejnowski, 1993), Monte Carlo (Brügmann, 1993), and Neural Networks (Richards, Moriarty, and Miikkulainen, 1998), were tested in Go. However, programs applying these techniques were not able to surpass the level of the best programs. The combination of search, heuristics, expert systems, and pattern recognition remained the winning methodology. Brügmann (1993) proposed to use Monte-Carlo evaluations as an alternative technique for Computer Go. His idea did not got many followers in the 1990s. In the following decade, Bouzy and Helmstetter (2003) and Bouzy (2006) combined Monte-Carlo evaluations and search in Indigo. The program won three bronze medals at the Olympiads of 2004, 2005, and 2006. Their pioneering research inspired the development of Monte-Carlo Tree Search (MCTS) (Coulom, 2006; Kocsis and Szepesvári, 2006; Chaslot et al., 2006a). Since 2007, MCTS programs are dominating the Computer Go field. MCTS will be explained in the next chapter. 2.6 Go Programs MANGO and MOGO In this subsection, we briefly describe the Go programs MANGO and MOGO that we use for the experiments in the thesis. Their performance in various tournaments is discussed as well.4