COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	\|\\M1BASL30\M2NGR40L\M3NGR25\M4NGR20\F2\CSTANFORD UNIVERSITY
C00010 ENDMK
C⊗;
\|\\M1BASL30;\M2NGR40L;\M3NGR25;\M4NGR20;\F2\CSTANFORD UNIVERSITY
\F3\CSTANFORD, CALIFORNIA 94305
\F4ARTIFICIAL INTELLIGENCE LABORATORY\←L\-R\/'7;\+R\→.\→S   Telephone:
\←S\→.415-497-3330
\F1\CMay 10 1977


Mr. Eric Jensen
Graduate School of Business
Duke University
Durham, North Carolina 27706

Dear Eric:

\JI have several comments to make on last Saturday's games.

In the first place the Duke game is not at all a clear cut loss for White,
as I had concluded.  I have not analysed the situation completely but it
seems to me that White can capture a piece, restoring the piece balance,
and then White has the so-called MOVE.  I had my program start at this
point and play itself and, while it did not push forward as I would have
liked, it did end in a draw situation.  Since your program seems to be
better than mine, it may still be able to find a win.  Would you be
willing to continue this game after your last move?  This was 10 15 for
move 59, I believe.  It would certainly look better to resolve the
situation, if the game is to be published.

Now as to the Stanford game.

This game seems to be lost for Black after the exchange ending in Black's
move 17 of 1-10, even though my program reported its moves through move 15
of 6-10 as being book moves and move 17 was forced.  This sequence of
moves breaks the conventional King-Row-Guard (holding squares 1 and 3)
and leaves Black open to the attack which your program used.  The trouble
seems to have started about move 9 of 16-20 where it seems to me that 2-7
would be much better here.  When I have my program play both sides, after
forcing the 2-7 reply, Black actually has the better game.

I am beginning to suspect that all of Black's moves from about move 5 of
8-11 are questionable even though my program reported them as book moves.
I am not yet willing to try to offer this as an alibi until I have done
some more checking and in any case if I depend on book moves then it was
my responsibility to see that the book moves were correct.  Unfortunately,
I gave all of my checker books away some six or seven years ago, so I can
not go back to these to check.  Do you hae a Lee's Guide, for example, and
can you check on these alleged book moves?

You will recall that I was unable to make the last version of my program
work because the Signature Tables had gotten garbaged and that I therefore
went back to an earlier version.  I now realize that this meant that I
also went back to an earlier version of the stored book moves.  One of the
last things that I did before I quit working on my program was to have a
checker expert go over the book games and correct them.  He found all
sorts of errors, some were in the books themselves.  Many other errors
were introduced because at the time these books moves were originally
punched into cards (by key-punch operators at IBM around about 1954) I had
them punch up every thing.  I failed to note that many losing moves where
shown, of course properly annotated, but the girls were not told to omit
these and so they got stored along with the good moves.  Of course, many
of these errors had been picked up during play but not all of them, and
even the expert may have missed some.

It is going to take a little work for me to unscramble this difficulty.
The files that the program actually uses have the material stored in a
highly condensed form showing the board positions with the recommended
moves and with no references to the original sources.  When there are more
than one recommended moves, the learning program stores up to 4 of these
and the playing program does a random selection to determine the move to
make.  My original method of correcting these files was to modify the
source game listings and then rerun the learning program.  I never
bothered to write a program to interrogate and correct the files directly
and it now seems that I should do this.

I am sorry that I was not better prepared for this match, both in terms of
having my program in better shape and in terms of having the mechanics of
playing better in hand.  It has been a long time since I even looked at
the program and many details had escaped my memory.  In any event, the
match did show you that your program was on the right track and that it
could do a creditable job.  I would like the opportunity to again
challenge it after I have had a chance to do a little work on my
program.\.

\←L\→S\←R\-L\/'2;\+L\→L



Sincerely,




Arthur L. Samuel



\←S\→L ALS:pdp10