SCANnews

32 S C I E N T I F I C A M E R I C A N J U N E 2 0 0 7

MA

TT

MO

SS

MA

N

the length and weigh one quarter as much. Moreover, during the summer spawning sea-son, the trapped females grow fewer and smaller oocytes, the cells that develop into eggs. In the creek, the gonads balloon from 3 percent of body weight to almost 8 percent. In the hot spring, organs increase from 1 per-cent of body weight to 2 percent. They would grow even less without submerged skin to nibble, Bardakci concluded. Some 90 percent of visitors arrive in the summer, providing a nutritional supplement at the perfect time.

The fi sh come from the carp and min-now family, which is known for adaptabil-

ity, says Richard Londraville, a biologist at the University of Akron. He adds that those in the hot spring may eventually evolve into a separate species in a few thousand years.

Other fi sh survive in waters as hot as or hotter than this spring, which hovers near 34 degrees Celsius. None are widely known to feed on skin, which may explain why G. rufa are catching on elsewhere. A Chinese spa-building company, which claims to have in-vented the concept, says on its Web site that it trained its own doctor fi sh and built 10 spas in China, including in Beijing. Some of the Turkish fi sh were scooped up and enlisted in springs in Japan, where at several spas they are now also performing fi sh pedicures.

Matt Mossman is based in Istanbul.

Silicon SmackdownNEW GO ALGORITHM AIMS TO DEPOSE HUMANS BY K AREN A. FRENKEL

A decade ago IBM’s chess program, Deep Blue, beat world champion Gar-ry Kasparov in a six-game match. The

event marked a milestone, forcing humans to yield dominance of yet another strategic diversion. Only the Asian board game Go seemed to be computer science’s Achilles’ heel: humans could soundly beat the ma-chines. A new algorithm can now take on strong human players—and win.

Go has proved enormously diffi cult for computer programmers because of the game’s deceptive complexity. The objective of Go is to stake out territory and surround

an opponent by placing black or white stones on the intersections of a nine-by-nine or 19-by-19 line grid. Especially on the large board, the number of possible moves per turn is huge—200 on average for each midgame po-sition compared with the several dozen pos-sible in chess. There are also enormous branching factors. Given N positions on the board, the total number of possible game positions is 3N, because every position can be occupied by a black or white piece, or it can be empty. The total number of legal po-sitions on the small board is about 1038; on the large board, about 10170. Additionally,

I I CO

MP

UTI

NG

BITE ME: Fish in hot-spring pools at a Turkish resort survive by feasting on human skin. The water does not sustain enough plankton and algae that would otherwise serve as food.

SCANnews

34 S C I E N T I F I C A M E R I C A N J U N E 2 0 0 7

more stones do not ensure victory, and play-ers must be able to consider local positions and the board as a whole.

To cope with such an enormous number of options, artificial-intelligence experts have designed algorithms to limit searches, but the programs have not been able to beat the better human players on large boards. Last fall two Hungarian researchers report-ed that their algorithm outdid the win rates of the best Go programs by 5 percent and could compete with professional Go players on small boards. Levente Kocsis of the Com-puter and Automation Research Institute at the Hungarian Academy of Sciences in Buda-pest and Csaba Szepesvári, now at the Uni-versity of Alberta in Edmonton, developed the algorithm, called UCT (for upper confi -dence bounds applied to trees). It extends the well-known Monte Carlo method.

First incorporated into Go programs in the 1970s, Monte Carlo works like a po-litical poll: it performs statistical sampling to predict the behavior or characteristics of a large group. When applied to Go, the al-gorithm evaluates and ranks candidate moves by playing a large number of random games. But playing the move with the high-est score in each position does not guaran-tee that the player will win the game. In-stead this type of search merely restricts

the number of relevant potential moves.UCT takes Monte Carlo further by fo-

cusing the search on the most promising moves. “The main idea is to sample actions selectively,” Kocsis says. The algorithm must strike a balance, testing alternatives that look the best at the moment to fi nd pos-sible weaknesses and exploring “less opti-mal-looking alternatives, to ensure that no good alternatives are missed because of ear-ly estimation errors,” he explains.

UCT calculates indices for moves and se-lects the move that has the highest index. The algorithm computes the index from the win rate, which describes how often that po-sition leads to a win, as well as the number of times the position has been visited but not played. UCT grows a decision tree in mem-ory and uses it to track these statistics. When UCT encounters a move that has not been visited previously, it adds the move to the tree and plays the rest of the game randomly.

UCT decides if the fi nished random game is a win or loss, then updates the statistics of all moves made during the game. If the index equals the win rate of the move, the algo-rithm quickly focuses on the most promising path. But nothing guarantees that an initial-ly successful path will eventually yield a win-ning move. So when selecting moves, UCT infl ates win rates by weight ing less visited candidate moves more heavily. The research-ers borrowed this idea from bandit prob-lems—selective weighting yields the maxi-mum gain for a gambler playing several slot machines with unknown average payoffs.

Mathematicians Sylvain Gelly of the University of Paris-South and Yizao Wang of the Polytechnic School outside Paris have incorporated UCT into a program they call MoGo. It has a 95 percent better win rate as compared with a previous state-of-the-art Monte-Carlo extension algorithm. Now the top-ranked Go algorithm, MoGo demon-strated its abilities this past spring, van-quishing strong amateur players on nine-by-nine boards and beating weaker ones on large boards. Gelly says that UCT is simple to implement and can be improved. So turn-ing the ultimate corner—ending the reign of professional human Go players—could oc-cur in 10 years, Kocsis states.

Karen A. Frenkel is based in New York City.

GO AT TACK : Humans still rule when it comes to Go, but a new algorithm can topple strong players.

GO BEYOND GAMES

The new Go-playing algorithm, called upper confi dence bounds

applied to trees (UCT), is not limited to games. It applies to any

problem that involves choosing the best option, as long as

alternatives have an internal treelike structure (that is, a

cascading set of choices) and their values can be recursively

computed. UCT may prove useful for targeting advertisements

on the Web, fi nding the best settings for an industrial plant

or optimizing channel allocation in cellular systems. S

HIN

JA

E W

ON

Th

e D

ail

y P

rin

ceto

nia

n