Language, Cognition, and Computation Lecture Series|
"Newcomb's Problem and Deterministic Choice: Implications for Cognitive Design"
Friday, May 28, 2004, 3:00 PM EST
Bartos Theatre, MIT Media Lab (E15)
AT&T Career Development Professor of Media Arts and Sciences
Cognitive Machines group
Newcomb's Problem, a long-standing decision-theoretic paradox, posits
an imaginary situation in which a simulator can reliably predict your
actions. A large reward was previously irrevocably set up for you if
and only if the simulator's prediction was that you would now make a
choice which (apart from the large reward) is slightly unfavorable to
you. A dilemma arises as to whether to make the choice that (almost
certainly) implies that you reap the large reward, even though the
choice does not cause your obtaining that reward.
The paradox bears on the compatibility of choice and determinism, a
quintessentially philosophical problem. But it also has important
ramifications for the science and engineering of cognition. Newcomb's
Problem illuminates a fundamental controversy about the foundations of
decision theory, which translates into a central question about the
design of intelligent, choice-making agents. What sort of relation
between contemplated action and goalcausal, subjunctive, evidential,
or some other relationmust hold for it to be sensible to consider the
action a means to the goal? That is, what constitutes a means-end
relation? And how might an agent that learns independently recognize
when this relation holds?
By appeal to relatively mundane, uncontroversial scenarios, Drescher argues
against evidential and causal criteria for linking means to ends, and
proposes instead a subjunctive criterion: an agent acts for the sake of
what would be the case if this or that action were taken, which (he argues)
is sometimes distinct from what an action either causes or gives
evidence of. Drescher sketches a computational implementation of subjunctive
means-end recognition, and shows how the proposed mechanism might
resolve Newcomb's Problem.
Gary Drescher received his PhD at the MIT AI Lab. His dissertation
proposed empirical-learning and concept-inventing machinery to account
for aspects of Piagetian cognitive development during infancy. Drescher
is the author of Made-Up Minds: A Constructivist Approach to Artificial
Intelligence. Now indulging in amateur philosophy, he is a visiting
fellow at the Center for Cognitive Studies at Tufts.
MIT Media Laboratory Home Page | Events Main Index