Like human players, the program had a large number of stored "patterns", and analyzing a position involved matching these patterns to suggest plans for attack or defense. By communicating plans down the tree, the analysis was verified and possibly corrected by a small search of the game tree (tens of positions) inluding specialized causality facility and quiescence search[1]. There were production rules to produce plans, implementing such concepts as checkmate, fork, skewer, and trapping the piece, etc.. A plan generator produced tactical plans in a Plan Language. The program is capable of finding very deep combinations because no limit is placed on its search depth. It searches for moves as long as a plan is continuing to work [2].
While Paradise was able to solve most of 92 positions picked from the first 100 from Win at Chess, with averaged three minutes thirty-three seconds for each solved position on a PDP-10[3], it was not able to play a complete reasonable game of chess due not further implemented knowledge employable in strategic, none-tactical positions, especially during the endgame. Controlling the search by recognizers, i.e. the amount to extend or to reduce if a move is accordant to a plan or not is still hot topic.
David Wilkins (1979). Using Patterns and Plans to Solve Problems and Control Search. Ph.D. thesis, Computer Science Dept, Stanford University, AI Lab Memo AIM-329
Tony Marsland (1987). Computer Chess Methods. Encyclopedia of Artificial Intelligence (ed. S. Shapiro). John Wiley & sons, New York. pdf draft, mentions Paradise on pp. 27
David Wilkins (1991). Working notes on Paradise chess patterns. Technical Note 509, AI Center, SRI International, 333 Ravenswood Ave., Menlo Park, CA 94025, pdf
a knowledge based chess program written at Stanford University in the late 70s by David Wilkins. Paradise was written in MacLisp, a dialect of the Lisp programming language developed at MIT within Project MAC. Paradise' goal was to find the best move in tactically sharp middlegame positions from the game of chess masters.
Like human players, the program had a large number of stored "patterns", and analyzing a position involved matching these patterns to suggest plans for attack or defense. By communicating plans down the tree, the analysis was verified and possibly corrected by a small search of the game tree (tens of positions) inluding specialized causality facility and quiescence search [1]. There were production rules to produce plans, implementing such concepts as checkmate, fork, skewer, and trapping the piece, etc.. A plan generator produced tactical plans in a Plan Language. The program is capable of finding very deep combinations because no limit is placed on its search depth. It searches for moves as long as a plan is continuing to work [2].
While Paradise was able to solve most of 92 positions picked from the first 100 from Win at Chess, with averaged three minutes thirty-three seconds for each solved position on a PDP-10 [3], it was not able to play a complete reasonable game of chess due not further implemented knowledge employable in strategic, none-tactical positions, especially during the endgame. Controlling the search by recognizers, i.e. the amount to extend or to reduce if a move is accordant to a plan or not is still hot topic.
Table of Contents
See also
Publications
[5]Forum Posts
External Links
Poogie Bell, Bobby Sparks, Keith Anderson, Patches Stewart, Grégoire Maret
References
What links here?
Up one Level