Cover Story / Ken Regan
speaker inside, similar to the way a chess engine appears to conceal a mini super- grandmaster. Searle argued that that when the English-speaking person (or a comput - er) follows a set of instructions to translate a language, no matter how well, they do not understand the language in the same way a native human speaker understands language. The same skepticism can be applied to whether or not computers un - der stand chess. Chess has often been described as a
form of language, and when I propose to Regan that today’s chess engines approach perfect play by following a set of rules embedded in source code, similar to the way the translator inside the Chinese Room follows a flowchart, he carefully considers his response. The implication that the best chess-playing entities on the planet follow rules revisits the ongoing debate in the chess community about whether or not human chess players also use rules to evaluate positions. Ironically, chess computers are commonly believed to play in a style that ignores rules. Regan speaks English, Spanish, Ger man,
Italian, and French, and he ap proach es the debate by distinguishing rules written in human language from those written in computer code. “When we get to the tunable parameters in the program,” says Regan, “all of the magic constants that define the value of the queen, the value of a rook, the value of a knight, the value of certain positional play, the values of squares, of attacks, these parameters are tuned by performance, linear regression. Program - mers don’t necessarily have a theory about what values or rules for those parameters work well. They have a general idea, but the final values are determined by [the engine] playing lots of very fast games against itself and seeing which values perform best.” When I insist that the ones and zeroes of an engine’s compiled code remain static, similar to rules written in a book, he leans back and restates his point. “Yes, that’s true. But computers use re - gression.” What Regan means by regression is this:
While some ones and zeroes remain static in the engine’s initial program, other ones and zeroes essential for the engine’s eval - uation function—those essential for the way it “thinks”—rapidly update in short-term random access memory (RAM). This process mimics training and en hances a computer’s ability to do more than just calculate. Regression creates real-time feedback that allows engines to “think” about each position unburdened by con text, similar to the way a human weighs imbalances. But computers calculate much faster. Deep Blue, the first computer to defeat a world champion in a standard time control match, succeeded despite its relatively poor evaluation function and made up for this deficiency via fast
28 June 2014 | Chess Life
calcula tion. Today’s top engines would destroy Deep Blue, because they evaluate better— because, ironically, they “think” more like a human.
“[ALAN] TURING WANTED TO MODEL human cognition with a computer, but I’m going in the opposite direction,” Regan says. “I want to use the computer to inform us about the human mind.” Regan’s data has reproduced a result in psychology first discovered by Nobel Prize-winning economist Daniel Kahneman and colleague Amos Tversky, which states that human perception of value is relative. “You’ll drive across town to save $4 on a $20 purchase, but you wouldn’t do it for a $2,000 pur chase,” says Regan. His data shows that players make 60 percent to 90 percent more errors when half a pawn ahead or behind than when the game is even. Regan claims that this is an actual cognitive effect, not a result of high-risk-high-reward play, because it is observed with players who have both the advantage and disad vantage. Chess has been called the drosophila (a small fruit fly, used extensively in genetic
research because of its large chromosomes, numerous varieties, and rapid rate of reproduction) of artificial intelligence. It is a popular re source for research in cog nitive science and psychology, because the Elo rating system provides an objective measure of human skill. Regan’s work follows this scientific tradition. He has processed over 200,000 reference games played by players ranging in Elo from 1600 to 2800, using Rybka 3 at depth 13 in single-line mode. Single-line mode is a bit less accurate than multi-line mode, but it runs roughly 20 times faster. These reference games provide a
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50 |
Page 51 |
Page 52 |
Page 53 |
Page 54 |
Page 55 |
Page 56 |
Page 57 |
Page 58 |
Page 59 |
Page 60 |
Page 61 |
Page 62 |
Page 63 |
Page 64 |
Page 65 |
Page 66 |
Page 67 |
Page 68 |
Page 69 |
Page 70 |
Page 71 |
Page 72 |
Page 73 |
Page 74 |
Page 75 |
Page 76