Letting Stones Go Unturned

Wéiqí, Baduk or simply, Go. Over two thousand years old, this challenging game of strategy and tactics has taught warriors, monks and intellectuals to focus, concentrate and plan. Far more complex than chess, there are more possible games than there are atoms in the known universe. Computers can now equal the performance of chess players, but none can equal us at this deceptively simple game. But one program is beginning to change our views of what is possible. In computer competitions this multi-armed bandit has stolen first prize more often than any other. It’s MoGo, or Monte Carlo Go.

15The game of Go may have simple rules, but to play successfully requires enormous skill. Few computer programs are able to navigate through the unimaginable number of possible moves and find the best strategy. The solution, as shown by MoGo, is not even to attempt to look at all the possibilities. Instead, a randomised exploration of the options based on multi-arm bandits makes this problem more manageable. The secret behind the early success of MoGo was Upper Confidence Trees.

The idea of using machines to take the role of our opponent in a game arose even before computers had been created. Chess has long attracted the attention of designers of these machines. While the very earliest chess automata were frauds, by the 1950s the subject was being considered seriously. According to Claude Shannon in 1950, we should think of the game as an imaginary tree branching out from the current state of the board. Each possible move was another branch from the current state; follow a branch and you reach a new board state from which new branches grow. Chess has a lot pieces and a lot of possible moves by each piece, making the game tree vast. Even the most powerful supercomputers cannot search all of this tree. However by the late 1990s and early 2000s, our best computers were fast enough to search significant parts of the tree. Computers such as Deep Blue and later Deep Fritz were able to beat human chess grand masters using these brute-force approaches.

But the game of Go was another matter entirely. Compared to chess, which may have around 37 legal moves possible each turn on average, Go has rarely fewer than 50 moves possible, and typically over 200 each turn. This means that the game tree for Go is considerably larger than the one for chess. So even with the very fastest supercomputers, a brute-force search for good moves simply doesn’t work for the game of Go.

Monte Carlo search is one useful way to make vast search spaces more manageable. Instead of trying to exhaustively search the whole space, a Monte Carlo method randomly samples points in the space and then uses the results of that sampling to approximate the solution. It’s like the game of Battleship, where random sampling is used to build an approximate picture of where the ships are on the grid. When applying Monte Carlo search to a game tree, this equates to evaluating many branches of the tree chosen randomly to see which player might eventually win or lose if those moves were followed, and then deciding on the best branch to follow. Traditional Monte Carlo methods might iteratively search the game tree, deeper and deeper, but they sampled each branch with a uniform probability. An even better approach, as introduced by researchers Levente Kocsis and Csaba Szepesvári  in 2006, is to bias the random sampling of branches.

Using a method called Upper Confidence Trees (UCT) (an extension of a previous method known as Upper Confidence Bounds) the different moves within the game tree are treated like different one-armed-bandit machines. Each time a one-armed-bandit is played it will return a random result, but on average some provide better results than others. In exactly the same way, each time a certain move is tried, it might result in a different outcome (depending on all the other moves), but on average some are better than others. By finding an appropriate balance between exploring which are the best bandits (moves) and exploiting the good ones already found, the Monte Carlo method can be biased towards more useful moves in the game tree. In this way UCT is able to search the game tree of Go selectively and randomly and discover a good move to make each time. (In reality “multi-arm bandits” are more appropriate for Go than one-arm bandits.)

In 2006 a group of researchers took exactly this approach when they developed their Computer Go playing program. Prof Remi Munos and his student Yizao Wang (a keen player of Go) at Ecole Polytechnique were inspired by the recent work of Rémi Coulom who had created a hugely successful Monte Carlo Go player called Crazy Stone. They joined forces with another student, Pierre-Arnaud Coquelin, and Olivier Teytaud and his PhD student Sylvain Gelly at University of Paris-Sud. This team applied the UCB algorithm to the game tree of Go and without realising it, independently invented the UCT algorithm at almost the same time as Kocsis and Szepesvári.

MoGo used UCT to bias its Monte Carlo search – a strategy that enabled the program to win the CGOS tournament where games were played on a smaller 9×9 grid, in August 2006.  MoGo was first introduced in the On-line Trading of Exploration and Exploitation Workshop 2006 (sponsored by PASCAL, the Network of Excellence on Pattern Analysis, Statistical Modelling and Computational Learning).  The same team went on to use a Monte Carlo and bandit-based  approach and win the 2006 “PASCAL Exploration vs. Exploitation Challenge,” set by TouchClarity (now Omniture TouchClarity), see box.

Despite the successes of the MoGo team, other researchers remained sceptical about their approach. According to Sylvain Gelly, “at that time, the ideas we were using in MoGo were considered ‘interesting’ in the Computer Go community only for the 9×9 Go, but were considered doomed by the fact that they were not successful on the “real” game. 9×9 Go was considered to be another game unrelated to 19×19 Go.”

But the MoGo team did not lose hope and Sylvain Gelly spent the next year improving MoGo (with methods such as Rapid Action Value Estimation). By March 2007 MoGo won a 19×19 Computer Go tournament against strong classical 19×19 programs, and was honourably ranked in 19×19 games against humans, which attracted interest outside the Computer Go community for these new techniques. MoGo continued to defy expectations, as Gelly explained: “Interestingly, the improvements made for the 19×19 board turned out to be useful for the 9×9 board as well (while people were claiming that the two games were unrelated), and MoGo in the mean time achieved “master” level (amateur) in 9×9, which also attracted interest outside Computer Go.”

Before long, MoGo was beating all other programs it played in the Kiseido Go Server tournaments, for several months in a row. Since then, MoGo has had many success, including winning the Gold Medal in the 2007 Computer Game Olympiad for the game of Go, and being the first program to defeat a professional player in 9×9 Go. In March 2008 MoGo won a 9×9 game against the winner of the “Tournoi de Paris”, the most important tournament in Europe, and Catalin Taranu (fifth Dan Pro) claimed that mogo was close to the Dan level in 19×19.

According to Olivier Teytaud, one of the current developers of MoGo, “MoGo does not contain a lot of Go expertise and is however quite strong; this shows that the approach is quite general… We are now applying UCT and related tools to other games and to some combinatorial optimisation problems.”

The idea of bandit algorithms used recursively for performing efficient tree search is a new research direction that is definitely worth investigating both from theoretical and practical perspectives.

From game-playing to webpage optimisation, Monte Carlo bandit methods seem to provide a powerful new technique. In the words of Remi Munos, “I think the idea of bandit algorithms used recursively for performing efficient tree search (such as UCT) is a new research direction that is definitely worth investigating both from theoretical and practical perspectives.”

TouchClarity was recently acquired by Omniture, specialists in online business optimisation. The development of MoGo has also been a fruitful avenue of research for all involved. In 2007 Yizao Wang and Sylvain Gelly won the prize for best student paper in the IEEE Symposium on Computational Intelligence and Games. Wang also received an honor “prix d’option” from Ecole Polytechnique for his work on this project. Gelly was awarded his PhD on “A Contribution to Reinforcement Learning; Application to Computer Go” in September 2007 and now works at Google, Zurich. Coulom continues to improve Crazy Stone which after several recent wins against MoGo is now perhaps the main competitor to MoGo in Computer Go.

Remi Munos summed up the MoGo work with enthusiasm: “From my point of view, this project (or adventure…) has been really a wonderful collaborative work which has led to some advances not only in the field of Go, but also which has opened new perspectives in other fields as well, such as in optimal control, large scale optimisation problems, and complex sequential decision problems, in particular when several agents interact, such as in games.”