Larry,

      Our conversation brought back many happy memories of my two years in Switzerland. Gerhard Rudolf and Werner Joho were wonderful colleagues who taught me a lot about computer programming which served me well over my physics career.

      Yes, we were really artificial intelligence amateurs playing around with a wonderful toy, the CDC 1604 computer. In fact the chess program's name stood for, "Chess Heuristics and Algorithms for Relaxing Lazy Yodelers", CHARLY. The program we used to play against Mac Hack 6 had been worked on in our spare time for about 1 year. The library openings were programmed by Gerhard Rudolf, our machine language expert, but the heart of the chess playing algorithm was programmed in Fortran. The library entries were collected from books, mainly, the 10th edition of Larry Evans, Modern Chess Openings. The moves were selected and tested by Werner Joho and myself, both of whom were merely recreational chess players.

      The main tree search went 4 plies deep, with the final position including an extensive evaluation of all possible exchanges on the square occupied by the piece being moved there on the last ply. This made the program relatively strong tactically but there was not much in the program incorporating strategy. Strategy, besides material advantage, included mobility of the pieces, pawn structure and advancement to queening. I do not recall much else about the program at this time.

      I do remember that it required us a considerable amount of time making the program totally "move-legal", such as incorporating all of the rules associated with "castling", "en passant", and "stalemate through move repetition". The only thing we did not include was pawn promotion to something other than a queen. I still can't think of a clever way to do that properly without making major modifications in the plausible move generator and the I/O move entry algorithm. I would love to hear from someone who has accomplished this in a satisfactory way.

      There is a really good article that was published in the 1967 Fall Joint Compute Conference, entitled, "The Greenblatt chess program". The authors listed are, Richard D. Greenblatt, Donald E. Eastlake, III, and Stephen D. Crocker. It is in the American Federation of Information Processing Societies Vol. 31 pages 801-810. I also have about 7 letters sent from MIT in our correspondence concerning how to set up the match.

      One of the more interesting problems of artificial intelligence that I learned about in working on the chess program was called, "The Horizon Effect". If a problem could be pushed "beyond the preprogrammed" tree search limit, the program would make the "bad" choice of sacrificing material to avoid losing, say, a queen "within" the horizon in the situation where the queen was going to be lost eventually anyway. I see this as not just a problem of "artificial" intelligence but human thinking in general, especially of politicians. You can find wonderful examples of this in today's headlines. For example politicians "buy" the favor of a certain class of workers by offering them fabulous retirement benefits. That effectively pushes an "accounting negative balance" beyond the horizon of the politician's career. The politician gets the immediate benefit of support from the affected workers but society will eventually have to pay the bill after the politician is gone.

Sincerely,

Gerald Tripard.

Richard Greenblatt also noted the "Horizon Effect", of course, although I don't know if he called it by that name. One way that Mac Hack VI attempted to combat it was via a mechanism that I believe was called "crossovers". If at the conclusion of the program's lookahead (typically six plies deep) there were still pieces under attack, the program extended the lookahead, for those positions only, to an even greater depth. This introduced some other problems, such as comparing board evaluations of different depths. However, this seemed better than working with a board evaluation based on a position that was still in a state of flux.