The Duke vs Stanford Computer-to-Computer Checker Match
		(Comments by Arthur Samuel)

A computer-to-computer checker match of two simultaneous games was played on
Saturday May 7 1977 beween a checker playing program recently written by Eric
Jensen and Tom Truscott of Duke University and the much older program written by
Arthur Samuel of Stanford University.  The two machines were not directly
connected and the moves as reported by each machine were relayed by phone and
entered into the other machine manually.  The games are appended.

When the match was terminated because of time constraints it was believed that
the Duke program had a possible win in both games.  Subsequent analysis reveals
that one of these games might still lead to a draw.  The other game was marred
by a bad Black move in the early part of the game.  The contest was therefore
not as definitive as one might have liked and a re-match seems indicated.  The
American Checker Federation has been written for an authoritative evaluation of
the games.

Some information about these two programs may be of interest.

The Duke program is a modern program especially planned to play a good game in
as little as 10 seconds of CPU time per move.  It was adjusted to take
approximately 25 seconds per move, and it was run on an IBM 370/165.  The
Stanford program was a reload of a 1970 version of the Samuel program,
esentially as it was described in Samuel's 1967 paper*.  This program had been
adjusted to take five minutes per move on the then existing Stanford machine.
No changes were made to this program other than those needed to get it to run on
the current Stanford PDP-10 computer under its present operating system and to
reduce the operating time in an arbitrary fashion to what turned out to be
approximately 22 seconds per move.

The Duke program has been design from the viewpoint of making it the very best
computer-playing program that can currently be produced by using as much
man-learned information about the game of checkers as possible.  The program is
a very creditable piece of work.  The Stanford program, by way of contrast, is a
machine-learning program that is given a list of possibly useful properties of
board positions and that is then charged with the entire task of adjusting the
weights assigned to these various properties on the basis of result obtained in
playing over games taken from published master play.  Both programs do a
variable depth tree search and both make extensive use of Alpha-Beta pruning.

The Duke program is, of course, entirely on its own once play commences, and it
must compute each move over again even when a previously seen board position is
encountered.  The Stanford program is always on its own in that it has never
been given any information as to the relative importance of the different
conditions that it is apt to encounter except as it can derive this information
by generalizing the situations encountered in playing against the book.  In
other words, the Duke program was given certain general principles while the
Stanford program was only given examples of good play.

The Stanford program does have one advantage in that it not only derives some
generalized principles from its learning runs but it is also able to make a
stored catalogue of the board positions that it has studied and it is able to
retrieve the recommended moves from this catalogue.  This feature results in a
substantial saving of time in the first few moves of a game when neither side
has deviated from acceptable book play, but it can cause trouble if any errors
are made in entering the book moves.  The generalized learning is reasonably
safe in that erronious results are more or less swamped out by the larger number
of correct moves that are analysed, but if a mistake has been made, this mistake
is always there to plague one should this exact board position be encountered
in subsequent play.  This seems to have happened in one of these games.

Mr. Jensen and Mr. Truscott deserve a great deal of credit for their work.  It
is too early, however, to attempt to draw any firm conclusions about the
relative value of the two quite different approaches.  Neither program has yet
been able to beat the very best human checker players.

The writer is very pleased by this renewed interest in programming checkers and
is looking forward to the time when a computer program will beat all comers.  He
still believes that the some form of machine learning will have to be used
before this goal can be reached.

- - - - - - - - - - - - - - - - - - - - - - - - - - 
*Some Studies in Machine Learning Using the Game of Checkers. II.
IBM Journal of Research and Development Vol. 11 No. 6 Nov.1967 pp 601-617.
- - - - - - - - - - - - - - - - - - - - - - - - - -

Appendix A
		Game S	Stanford playing Black
	11 16	11 16	3 7	16 19	19 24	20 24
	22 18	19 15	18 15	23 16	7 2	23 18
	10 14	10 19	9 13	14 23	9 13	24 27
	24 19	24 15	15 6	27 18	18 14 	32 23
	8 11	6 10	2 9	12 19	24 28	28 23
	25 22	15 6	29 25	21 14	25 22	11 7
	7 10	1 10	8 11	11 16	16 19	32 27
	27 24	28 24	22 18	15 11	26 23	7 3
	16 20	4 8	13 17	7 10	19 26	27 32
	31 27	24 19	19 15	14 7	30 23

		Game D	 Duke playing Black
	12 16	4 8	5 9	10 15	15 18	25 30
	24 20	25 22	30 25	28 24	25 21	9 6
	11 15	16 20	9 14	15 18	18 22	1 10
	20 11	22 17	27 24	14 9	17 14	14 9
	7 16	9 14	20 27	18 23	3 8	8 11
	22 18	18 9	31 24	17 13	9 5	5 1
	15 22	6 22	14 18	8 11	23 26	30 25
	25 18	26 17	19 16	21 17	21 17	9 6
	8 11	2 7	11 27	11 15	22 25	10 15
	29 25	23 19	32 14	24 20	13 9	6 2