Chess on the IBM 704, full board
In May 1958, at the Western Joint Computer Conference, Alex Bernstein, Michael Roberts, Timothy Arbuckle and Martin Belsky described a chess program for the IBM 704 that plays a complete game on the 8 × 8 board, with castling, promotion and en passant. It considers at most seven «plausible» moves per position, looks four half-moves ahead and thinks eight minutes a move on average.
Why it matters
The program brings the selection of moves into chess programming: of about 800,000 possible continuations two moves deep it examines about 2,500. That count is by Newell, Shaw and Simon in September 1958, not by the authors. Its play is uneven: in the one full game the authors of the survey knew of, it lost to a good player. This is an editorial assessment.
What the authors' paper says. Five parts: input-output, table generation, evaluation, decision, tree. Evaluation from four criteria: mobility (the number of available moves), area control, king defence, material; the first three are added, material is multiplied by a large factor, so the program does not sacrifice material and likes to exchange when ahead; the ratio of the two sides' scores is taken. The Plausible Move Table is limited by «time and coding considerations» to seven. The tree: with seven moves and four half-moves of depth the examination takes 8 minutes on average; a selective mode cuts branches whose score is below the position's score before the move and brings the average to 2 minutes, but play becomes more conservative. The authors write that the program plays a «passable amateur game» in the opening and middle game and is weak toward the end. In the discussion Bernstein says the machine has so far played white only and is being debugged at MIT. What the survey by Newell, Shaw and Simon adds (RAND P-1319, September 1958). About two years of work, six months in partial operation; one full game played, in which a good player beat it; of about 800,000 legal continuations two moves deep it examines 2,500; memory of 7,000 words against 600 for the Los Alamos program; a move in 8 minutes against 12 for Los Alamos (on a different, smaller 6 × 6 board and a different machine). An example of a blind spot: a bishop is attacked by three pawns in a row and the program does not see it, because no generator proposes the corresponding moves. What the record does not claim: that this is the first chess program or the first to play a complete game (the word «first» is not in the paper; it is the claim of antiquarian listings, which were not checked); results against people beyond those given; the content of the Scientific American article, which was not read.