To: Eric  Bloedorn,  Ibrahim  Imam,  Ken  Kaufman,  Mark  Maloof,
Ryszard  Michalski,   and  Janusz  Wnek,   of  The  George  Mason
University East-West Challenge Group

From:  Donald Michie

5 Oct

Dear colleagues

David showed me your extremely  interesting  message.  Both  your
points  are  in  our  view valid. Indeed I see one of them, to do
with the discrepancy between "P-complexity"  and  true  cognitive
complexity,  as central to the whole ML enterprise. How nice when
we can replace Minimal Description Length with Minimal  Cognitive
Load as a driving principle in our ML algorithms!

Meanwhile we used the best we had as a measuring stick. I discuss
the  issue  towards the end of my article on Competition 1, which
will appear in Computing on 20 October. The full text is  in  the
file  michie.txt,  and gives details of the part of this competi-
tion that was attacked by unaided human brains,  42  of  them  in
this case, after subtracting frivolous entries. It also announces
that for Computing readers the present competition is  to  become
the first of a regular series, held every two months (see closing
passage of accompanying document). This  interest  shown  by  the
editors  of  Computing  came  as  a  pleasant  surprise. Overseas
readers can enter by email. To get the text of my articles alter-
nately  setting  and  assessing  successive  Trains  competitions
you'll need to get onto Computing's mailing list (their email  is
"computed@cix.compulink.co.uk".  I'll keep you posted with my own
email address.

In fairness to Bernhard Pfahringer, I should add  my  view  that,
given the game rules, his was an exceptionally impressive effort,
well deserving of the prize. As stated in the accompanying  docu-
ment,  for future competitions I shall base complexity scoring on
Trenglish, which has now, I believe, evolved into a  sufficiently
mature  and versatile formal language to take the strain. This is
essentially Janusz  Wnek's  "  ...  number  of  words  in  simply
expressed  and grammatically correct English sentences represent-
ing condition parts of the rules".

On your second point concerning Competition 2, you appear to have
proved a (to us) unwelcome theorem, to the effect that anyone who
has submitted to Competition 1 a rule that scores  in  the  lower
quartile of Trolog scores can automatically secure full marks for
Competition 2 simply by declaring in his/her  submission  of  his
50-train labelling that he/she wishes the said rule to be used as
oracle in assessing the labelling. Having actually used this rule
to assign the labels, he will be assured of a 100% correct score!
More than one reader used this tactic, so a tie-break was applied
(see  Ashwin  Srinivasan's  note  in  the  accompanying materials
briefly reporting on Competitions 1, 2 and 3). The cause  of  our
slip-up  was  an  unexamined  assumption that the intersection of
Competition 1 entrants and Competition 2 entrants would be empty.
The assumption was of course unwarranted.

David Page is circulating this message and its  appended  article
to AI Comp and ML List, together with Ashwin's report on Competi-
tions 1, 2 and 3.

Donald Michie.  6 Oct 94
