290 DONALD MICHIE
box for the permutation of O top left
and X bottom right did not contain
beads for putting the next O on
those squares. Michie considered
that putting beads for all nine
possible O positions in every box
would “complicate the problem
unnecessarily.” It meant MENACE
would not only learn to win or draw,
it would also have to learn the rules
as it went along. Such start-up
conditions might lead to one or two
early disasters that collapsed the
whole system. This demonstrated
a principle: machine learning works
best starting simple and gradually
add more sophistication.
Michie also pointed out that
when MENACE lost, its last move
was the 100 percent fatal one. The
move before contributed to the loss,
as though backing the machine into
a corner, but less so—usually it still
left open the possibility of escaping
defeat. Working back toward the
start of the game, each earlier move
contributed less to the final defeat—
that is, as moves accumulate, the
probability that each becomes the
final one increases. Therefore as
the total number of moves grows, it
becomes more important to get rid
of choices that have proved fatal.
Michie simulated this by having
different numbers of beads for
each move. So for MENACE’s
second move (third move overall),
each box that could be called upon
to play—those with permutations
of one O and one X already in the
grid—had three of each kind of
bead. For MENACE’s third move,
there were two beads of each kind,
and for its fourth (seventh move
overall), just one. A fatal choice
on the fourth move would result
in removal of the only bead
specifying that position on the
grid. Without that bead, the same
situation could not recur.Human vs MENACE
So what were the results? Michie
was MENACE’s first opponent
in a tournament of 220 games.
MENACE began shakily but soon
settled down to draw more often,
then notch up some wins. To
counter, Michie began to stray
from safe options and employ
unusual strategies. MENACE took
time to adapt but then began
to cope with these too, coming
back to achieve more draws, then
wins. At one point in a series of
10 games, Michie lost eight.
MENACE provided a simple
example of machine learning
and how altering variables could
affect the outcome. Michie’s
description of MENACE was, in
fact, part of a longer account that
went on to compare its performance
with trial-and-error animal learning,
as Michie explained:
‘“Essentially, the animal makes
more-or-less random movements
and selects, in the sense that it
subsequently repeats, those which
produced the ‘desired’ result. This
description seems tailor-made
for the matchbox model. Indeed,
MENACE constitutes a model of
trial-and-error learning in so pureExpert knowledge is intuitive;
it is not necessarily accessible
to the expert himself.
Donald MichieDonald Michie Born in 1923 in Rangoon, Burma
(Myanmar), Michie won a
scholarship to Oxford in 1942, but
instead assisted in the war effort
by joining the code-breaking
teams at Bletchley Park, becoming
a close colleague of the computing
pioneer Alan Turing.
In 1946, he returned to Oxford
to study mammalian genetics.
However, he had a growing
interest in artificial intelligence,
and by the 1960s it had become
his main pursuit. He moved to the
University of Edinburgh in 1967,
and became the first Chairman
of the Department of MachineIntelligence and Perception.
He worked on the FREDDY
series of visually-enabled,
teachable research robots. In
addition, he ran a series of
prestigious artificial intelligence
projects and founded the
Turing Institute in Glasgow.
Michie continued as an
active researcher into his
eighties. He died in a car
accident while traveling
to London in 2007.Key work1961 Trial and Error