Optimization by parameter-perturbation correlation

Abstract

This paper describes a simple method of steepest-descent optimization of a criterion functional F(α₁, α₂, ^...) depend ing on the state variables y _i(t;α₁,α₂,^...) of a dynamic system with design parameters α ₁ , α ₂ ^.... The system is simulated on a fast analog computer and the parameters are given mutually orthogonal sequences of binary pertur bations during successive computer runs. Simple correla tion of each parameter perturbation with the criterion functional value at the end of each computer run yields ap proximations for the gradient components ∂F/∂α _k needed for steepest-descent optimization with a minimum- of dig ital logic. As an example, two- and three-parameter model- matching problems are solved by iterative computation at 1000 iterations per second.

A course term-paper for first-year graduate students