A chess program built from goals, not numbers
On 4 September 1958 Newell, Shaw and Simon issued a RAND preprint on chess programs (the IBM Journal printed it in October). Having compared the programs of Shannon, Turing, Los Alamos and Bernstein, they describe their own: a set of independent «goals», each of which proposes moves and evaluates a position, with moves compared not by a sum of numbers but by the order of the goals. The program is coded in IPL-IV for JOHNNIAC, and in the spring of 1958 only two goals were working.
Why it matters
The record holds two things together: a comparative history of chess programs written by competing authors, and the description of a program that deliberately gave up a numerical additive evaluation because, in the authors' view, people do not think that way either. About the play of the program itself the paper says almost nothing: it is not debugged. This is an editorial assessment.
The design. The goals are independent modules (king safety, material, centre control, development, king-side attack, promotion, pawn structure); each has a move generator, a static evaluation and an analysis generator. The value of a move is a vector whose components depend on the list of goals in that position; values are compared lexicographically, so a goal of higher priority wholly dominates. Move choice: the first acceptable move by a threshold, and if there is none, the best found. Status in September 1958. The program «was coded this spring and is not yet fully debugged»; two goals are coded, material and centre; development and pawn structure are defined; king safety, serious threats and gambits are still to be added. It is written in IPL-IV, which an interpreter runs on JOHNNIAC. The size now is 6,000 words, 16,000 expected, of which about 1,000 words of machine code and 3,000 of IPL; the estimated time is from one to ten hours a move (a few minutes in quiet opening positions), «at least ten times slower» than Bernstein's. The authors do extensive hand simulation on typical positions. What the paper says about others (from their publications): the Los Alamos program of 1956 on MANIAC I played on a 6 × 6 board without bishops, about 12 minutes a move; Bernstein, seven moves per position, 2,500 positions of about 800,000; Turing counted material in «dead» positions. What the record does not claim: that the program played even one game (the table of results is garbled in the scan and was not used); that the printed version in the IBM Journal equals the preprint word for word (it was matched only by title and first sentence); the later fate of the program. This is the authors' own description, hence the medium confidence.