Computer Chess

Computer chess is the long effort to make machines play chess, which became artificial intelligence's most visible benchmark for roughly half a century: from paper designs in the late 1940s, through programs on room-sized mainframes, to dedicated hardware that beat the world champion in 1997, and finally to engines on a phone that no human can hope to match.

The intellectual blueprint was Claude Shannon's 1950 paper Programming a Computer for Playing Chess, published in Philosophical Magazine, which set out the two strategies that framed the field for decades. Type A programs search all legal moves to a fixed depth and score the frontier with an evaluation function; Type B programs search selectively, pursuing only plausible lines as human players do. Shannon also estimated that the game tree is astronomically large, which is why a brute-force enumeration to checkmate was ruled out and why evaluation and search had to cooperate. Alan Turing was working along parallel lines: with David Champernowne he wrote a paper program, usually called Turochamp, in the late 1940s, and since no suitable computer was at hand, Turing executed its rules by hand, move by move, in a game against a human opponent. The first fully working programs followed in the mid-1950s. The Los Alamos chess program of 1956, run on the MANIAC I computer, played on a reduced six-by-six board without bishops to keep the search tractable, and became the first program to defeat a human, a beginner who had just been taught the rules.

The 1960s and 1970s turned theory into competition: Alan Kotok's program at MIT, developed under John McCarthy for the IBM 7090 and playing a correspondence match against a Soviet program from the Institute for Theoretical and Experimental Physics in 1966 and 1967, showed that alpha-beta pruning made deeper search practical. Annual computer chess championships began in 1970, and programs such as Chess 4.x from Northwestern University dominated them. Ken Thompson and Joe Condon at Bell Labs built Belle, which used special-purpose move-generation hardware, won the 1980 world computer championship, and in 1983 became the first machine to earn a US master rating. The common lesson was that, for chess, speed mattered more than imitating a master's judgment: each extra ply of search bought a measurable rating gain.

The milestone itself came in May 1997, when IBM's Deep Blue defeated the reigning world champion Garry Kasparov in a six-game match in New York by a score of 3.5 to 2.5. Kasparov had beaten an earlier version of the machine 4 to 2 in Philadelphia in February 1996. Deep Blue descended from Deep Thought, built by Feng-hsiung Hsu and Murray Campbell and colleagues first at Carnegie Mellon and then at IBM, and combined custom chess chips with a parallel IBM system, searching on the order of two hundred million positions per second according to its designers. In the 1997 match Kasparov won game 1, Deep Blue won game 2, games 3 through 5 were drawn, and Deep Blue won game 6. The event is often summarized as the machine out-thinking the man, but its heart was brute-force search with a large hand-built evaluation function, tuned with grandmaster help.

A frequent misconception is that Deep Blue showed machines had become generally intelligent, or that it learned the way people do; it was a specialist whose strength came from enumerating positions at enormous speed, and it could do nothing else. The 2017 contrast is instructive. Stockfish, an open-source engine that grew out of the Glaurung engine and was released in 2008, is still an alpha-beta searcher, later supplemented with a small neural-network evaluation. AlphaZero, described by David Silver and colleagues at DeepMind in a December 2017 preprint and published in Science in 2018, learned chess from the rules alone through self-play and evaluated far fewer positions per second; it won a 100-game match against Stockfish 8 without losing a game, although critics pointed to the conditions and the version used. Modern engines blend both lines of thought, and computer chess continues to matter as a testbed for search, evaluation learning and human-machine collaboration.

Read computer chess as the field where search, evaluation and raw hardware were tested against a clearly measurable opponent: Shannon drew the map, Deep Blue proved brute force could win, and AlphaZero showed that learned judgment could do it with a very different recipe.