Statistical Physics, Bounded
Rationality, and Distributed Control/Optimization
David Wolpert
Computational Sciences Division
NASA Ames Research Center
Wednesday, December 1, 2004
4:30 - 5:45 PM
Terman Engineering Center, Room 453
Abstract:
A major problem in optimization and control is how to implement control
on (massively) distributed systems, i.e., how to do distributed
optimization. Especially difficult is how to do this in an adaptive
manner, with mixed types of control variables. A long-running
difficulty with conventional game theory is how to modify it to
accommodate bounded rationality. Finally, a recurring issue in
statistical physics is how best to approximate joint probability
distributions with decoupled (and therefore far more tractable)
distributions.
This talk shows that the same information-theoretic structure, known as
Probability Collectives (PC), underpins all three issues. This
means that distributed control/optimization, statistical physics, and
game theory are fundamentally identical. Accordingly techniques
and insights from one of those fields can be applied to the
others. One example of this, presented here, is how to apply
steepest descent techniques to optimize/control systems of discrete
variables. Another example is the use of the grand canonical ensemble
of statistical physics to elaborate game theory in which the number of
players is not pre-determined, but varies stochastically.
Operations Research Colloquia: http://or.stanford.edu/oras_seminars.html