Feb 8, 2007

A new way to play computer-Go?

Computers have started to outperform humans in games they used to lose [full text here]

[...] Deep Blue and its successors beat Mr Kasparov using the “brute force” technique. Rather than search for the best move in a given position, as humans do, the computer considers all white's moves—even bad ones—and all black's possible replies, and all white's replies to those replies, and so on for, say, a dozen turns. The resulting map of possible moves has millions of branches. The computer combs through the possible outcomes and plays the one move that would give its opponent the fewest chances of winning.

Unfortunately, brute force will not work in Go. First, the game has many more possible positions than chess does. Second, the number of possible moves from a typical position in Go is about 200, compared with about a dozen in chess. Finally, evaluating a Go position is fiendishly difficult. The fastest programs can assess just 50 positions a second, compared with 500,000 in chess. Clearly, some sort of finesse is required.

In the past two decades researchers have explored several alternative strategies, from neural networks to general rules based on advice from expert players, with indifferent results. Now, however, programmers are making impressive gains with a technique known as the Monte Carlo method. This form of statistical sampling is hardly new: it was originally developed in the Manhattan project to build the first nuclear bombs in the 1940s. But it is proving effective. Given a position, a program using a Monte Carlo algorithm contemplates every move and plays a large number of random games to see what happens. If it wins in 80% of those games, the move is probably good. Otherwise, it keeps looking.

This may sound like a lot of effort but generating random games is the sort of thing computers excel at. In fact, Monte Carlo techniques are much faster than brute force. Moreover, two Hungarian computer scientists have recently added an elegant twist that allows the algorithm to focus on the most promising moves without sacrificing speed.

The result is a new generation of fast programs that play particularly well on small versions of the Go board. In the past few months Monte Carlo-based programs have dominated computer tournaments on nine- and 13-line grids. MoGo, a program developed by researchers from the University of Paris, has even beaten a couple of strong human players on the smaller of these boards—unthinkable a year ago. It is ranked 2,323rd in the world and in Europe's top 300. Although MoGo is still some way from competing on the full-size Go grid, humanity may ultimately have to accept defeat on yet another front.

Copyright © 2007 The Economist Newspaper and The Economist Group. All rights reserved.

No comments: