ADAPTIVE PLANS OPTIMAL FOR PAYOFF-ONLY ENVIRONMENTS.
Abstract
The paper characterizes a class of adaptive algorithms, the reproductive plans, which produce optimal performance in conditions where the information fed back to the algorithm consists only of a payoff at each instant of time. The payoff function can by any bounded (non-linear) function of the algorithm's output space. The reproductive plans have the advantage that they achieve a global optimum (over the whole time course) via a local step-by-step optimization of well-defined quantities. The theorem guaranteeing optimal performance is a modification of the Kuhn-Tucker fixed point theorem closely related to Gale's work in mathematical economics. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- May 01, 1969
- Accession Number
- AD0688839
Entities
People
- John H. Holland
Organizations
- University of Michigan