Searching nonlinear functions for high values (Q1123828)

One way of describing complex systems like adaptive nonlinear networks (ANN) is to represent the ANN's component structures (be them rules, strategies, chromosomes, or the like) as a collection of k-bit strings. The author is concerned with modelling the ANN's search as a sampling on the space of k-bit strings using a probability distribution p(t) - which changes as time t increases. Each k-bit x represents a structure to be tried and a real-valued function u(x) can be helpful in biasing the distribution p(t) to direct the search. The idea is to ``re-represent'' the information given by u using a hyperplane transformation. The problem then is to design a feasible algorithm that (as information accumulates) provides the biases suggested by the hyperplane transform. In this respect, the author shows that genetic algorithms [see e.g. \textit{J. H. Holland} et al., Induction: Processes of inference, learning and discovery. MIT Press (1986), \textit{J. J. Grefenstette}, Genetic algorithms and their applications. (1987)], viewed as hyperplane-directed search procedures, rapidly provide the biasing implied by the hyperplane transform without explicitly carrying out the calculations involved.