Adaptive automata based on Darwinian selection

The results are: 1) The continuity of desired dynamical properties differs in the two ensembles, and 2) limitations in the capacity of mutation selection procedures, and approaches to overcoming them.

Abstract

The principle of natural selection is general, and it is natural to assess its implications for achieving automata with desired dynamical or structural properties. The following issues arise: 1) Appropriate definition of the analogue of "genotype" and "phenotype". 2) Definition of the ensemble of "possible" automata in which mutational search for desired properties is occurring. 3) Kinematic properties of the "fitness landscape" in the ensemble governing the statistical features of connected walks through fitter variants. 4) Optimal mutation selection strategies given a particular fitness landscape. We assess these questions in two ensembles of automata under selection for one attractor which matches a predetermined "target pattern". The results are: 1) The continuity of desired dynamical properties differs in the two ensembles. 2) When the best automaton seeds each generation, selection follows a characteristic curve, and asymptotes at automata which approach but fail to achieve the desired attractors. 3) Designation of a subset of the variables as hidden from fitness estimation as part of the target pattern does not improve approach to the target pattern. We discuss limitations in the capacity of mutation selection procedures, and approaches to overcoming them.