Agoric/Genetic Methods
in Stochastic Design
Rutgers University
This is an abstract
for a talk to be given at the
Fifth
Foresight Conference on Molecular Nanotechnology.
The full paper is available at
http://www.cs.rutgers.edu/~josh/chsmith.html
Molecular nanotechnology will be physically capable of
producing objects whose complexity exceeds that of any currently
designed artifact by several orders of magnitude. Designs which
make more than non-trivial use of this capability are beyond the
capabilities of human designers, even using current CAD methods.
Thus automatic design, or at least an improvement in automation
over existing practice, will be a crucial component of molecular
manufacturing.
Central problems in automatic design include the allocation of
scarce resources in the design (e.g. power and materials
budgets), managing tradeoffs between conflicting design goals,
and control of the overall design process itself and the
simulations that it entails. The present effort is an
investigation of a mixed-paradigm control model, drawing from
evolution (the "genetic algorithm") and economics
("agoric algorithms"). We show that this model is a
promising formulation for the general control and integration
task. We present experimental results in which it performs
certain desirable control tasks, including rational allocation of
effort in stochastic methods, coordinating local expertise into
an overall structure using the price mechanism, and driving the
overall process towards global obtima.
*Corresponding Address:
J. Storrs Hall, Deptartment of Comp. Science, Rutgers University,
New Brunswick NJ 08903, ph: 908-445-3896, fax: 908-445-0537,
email: [email protected]
|