Genetic algorithm design of neural network wavefront predictors

Abstract

A genetic algorithm (GA) is employed to determine the structure of measured, Shack-Hartmann data and its optimum artificial neural-network (ANN) predictor. In the GA approach there are no preordained architectures imposed. The NN architecture that evolves out of many generations of adaptation can also be interpreted as a mapping of the signal complexity. The GA approach inherently addresses the problems of generalization, over fitting of data, and the trade-off between ANN complexity and performance. One objective was to establish how much improvement could ideally be expected from NNs compared to linear techniques. The principal conclusions are: (i) The main input-output relationship is linear with only a small contribution from the nonlinear elements. (ii) The improvement achievable with ANNs compared to optimal linear predictors was less than a 10% reduction in predictor error. (iii) The optimum temporal input window of tip-tilt data corresponds to the time constant introduced by aperture averaging.