Samuel's signature tables
In November 1967 Arthur Samuel published the sequel to his 1959 paper on the checkers program. In place of a linear polynomial of position features he describes a hierarchy of «signature tables» and an improved way of learning from master games. The correlation of the evaluation with the master's move rose to 0.48 after about 175,000 moves for the tables, while for the polynomial it levelled off at 0.26 after about 50,000.
Why it matters
The record gives a numerical answer to what changed in eight years: not the strength of play but the quality of the learning procedure. The author writes that the program still cannot outplay checker masters, though it plays much better. This is an editorial assessment.
What was read (as excerpts of a full-text search). The abstract: «a new signature table technique» and «an improved book learning procedure» which the author thinks much superior to the linear polynomial method described earlier. The program uses about 27 parameters (position features); the coefficients come from an analysis of about 150,000 «book» moves taken from master games; a sample of about 250,000 positions of master play was key punched and transcribed to magnetic tape. The measure of a feature's quality is C = (L - A)/(L + A), where L is the accumulated count of all available moves (the rest of the definition was not read); by the author it has the dimensions of a correlation coefficient. For the polynomial the coefficient «appears to stabilize at a figure of 0.26 after about 50,000 moves», and for the signature tables it «continues to rise and finally after perhaps 175,000 moves reaches a limit of 0.48». The author writes: «While still unable to outplay checker masters, the program's playing ability has been greatly improved». What the record does not claim: any figure of playing strength or results against people (not read); that the signature tables learn without human moves (the method rests on book moves); the details of the table structure (the diagrams and the sections between the excerpts were not read). The whole paper was read as excerpts and there is no second source, hence the medium confidence.