[gnu.chess] GNU Chess 1.55 ready

cracraft@AI.MIT.EDU (Stuart Cracraft) (01/02/90)

	[ This software will soon be available for anonymous uucp on
	  osu-cis and anonymous ftp on gatekeeper.dec.com. ]

Forwarded-For: cracraft@ai.mit.edu

Send bug reports to cracraft@ai.mit.edu.

Version 1.55 of GNU Chess is now available. This version has several
important new features and results. A compressed tar file of 1.55
lives in prep.ai.mit.edu:u/emacs/gnuchess-1.55.tar.Z.

NEW FEATURES

   * A faster move generator. This allows the program to search
     deeper and faster in the same amount of time.
   * Think-on-opponent's time is now available in glass-tty
     mode. This had been available in curses mode. To turn
     on the feature, ensure easy-mode is off by toggling
     with "easy" command and then checking with "help" command.
   * Certain bad lines have been removed from the opening
     book and other good ones added to patch opening blindness
     especially in play against match partner Mach 3 (see below).
   * Improved draw observance for detecting draws in the search
     tree. This permits the program to play more strongly since
     it sees threats of a draw more accurately.

MATCH RESULTS

A 10 game match was conducted between GNU Chess 1.55 running on a Sun
Sparcstation-1 and the strong commercial chess machine Fidelity Mach 3.
The Fidelity program is officially rated USCF 2265 (2200 is master). Most
observers acknowledge it is a true master. All previous matches between
GNU (or derivatives) and Mach 3 resulted in no more than 3 points
for GNU, sometimes only 1.5 points.

Startlingly, the match result ended in 7-3, GNU's favor! Using
statistical rating tables, this corresponds to a 146-153 point rating
advantage for GNU. Normally, if two people were playing with each
other, we'd just add this range to the loser's rating (or the average
rating of the opponents). If we did this, GNU's rating becomes:

   2265+146 to 2265+153	=>  2411 to 2418   (senior master)

However, since computer vs. computer testing tends to overstate
tactical advantage, a correction factor of 0.75 is generally
applied to the rating advantage range:

   2265+(146*0.75) to 2265+(153*0.75) => 2374 to 2380 (near senior master)

But most observers feel Fidelity Mach 3 was lucky to get its rating
of 2265 which they feel is really 2215 to 2220 (still master).
With this in mind, adding the correct rating advantage range 
to the corrected Fidelity rating range we get (for GNU's final
rating):

    2215+(146*0.75) to 2220+(153*0.75) => 2325 to 2335 (strong master)

This result is most unexpected since all other versions of GNU had
scored no more than 3 points out of 10 against the Mach 3. As far as
we can tell, the big leap came due to a couple of reasons: (1) the
move generator speeds up the program significantly and (2) the
Sparcstation-1 seems peculiarly suited to speedy chess processing.
(Also, minor modifications to the book, draw factor, and thinking on
opponent's time have helped considerably.)

Please remember this rating is based on a short match result.
Statistical confidence in the range endpoints is limited. Also,
certain moves GNU plays are clearly non-master in quality. Computer
masters generally achieve their strength through accuracy of tactics,
not subtle positional moves.

Caveat emptor.

  Stuart Cracraft
  Email: cracraft @ wheaties.ai.mit.edu
  Phone: (714) 770-8532