Monte-Carlo approximation of temperature

Summary

Monte-Carlo tree search is a powerful paradigm for the game of Go. We propose to use Monte-Carlo tree search to approximate the temperature of a game, using the mean result of the playouts. Experimental results on the sum of five 7x7 Go games show that it improves much on a global search algorithm.

Monte-Carlo Go has recently improved to compete with the best Go programs [Coulom 2007; Gelly et al. 2006; Gelly and Silver 2007]. We are interested in the use of Monte-Carlo methods when there are independent games. In such cases it might be interesting to analyze the games independently instead of considering them as a unified game.

Section 2 describes related works Section 3 presents the Monte-Carlo algorithms we have tested. Section 4 details experimental results. Section 5 concludes.

In this section we expose related works on Monte-Carlo Go. We first explain basic Monte-Carlo Go as implemented in Gobble in 1993. Then we address the combination of search and Monte-Carlo Go, followed by the UCT algorithm, and previous works on the approximation of temperature.

The first Monte-Carlo Go program is Gobble [Brueg-mann 1993]. It uses simulated annealing on a list of moves. The list is sorted by the mean score of the games where the move has been played. Moves in the list are switched with their neighbor with a probability dependent on the temperature. The moves are tried in the games in the order of the list. At the end, the temperature is set to zero for a small number of games. After all games have been played, the value of a move is the average score of the games it has been played in first. Gobble-like programs have a good global sense but lack of tactical knowledge. For example, they often play useless Ataris, or try to save captured strings.