The World of SPLectrum
Home > Positioning > Subjects > Evolution > Evolutionary computation
Evolutionary Computation
In computing, “evolution” names a method: a population of candidate solutions is varied, the variants are scored, and the better ones are kept and varied again. It is the home of the word where every ingredient of the Darwinian recipe, heredity included, is chosen by the designer; its nearest neighbours, breeders’ artificial selection and the directed evolution of molecules in the laboratory are built on purpose too but work with biological heredity.
Origins
Alan Turing’s “Computing Machinery and Intelligence” (1950) proposed educating a “child machine” by trial and improvement, and drew the comparison himself: “There is an obvious connection between this process and evolution”, with the structure of the child machine as the hereditary material, changes to it as mutation, and “judgment of the experimenter” in the place of natural selection. He hoped it would be faster: “The survival of the fittest is a slow method for measuring advantages.” Nils Aall Barricelli ran numerical simulations of evolving patterns at the Institute for Advanced Study in the early 1950s, and Alex Fraser simulated genetic systems later in the decade.
Four methods
Four programmes grew up largely independently in the 1960s. Lawrence Fogel’s evolutionary programming, set out with Owens and Walsh in Artificial Intelligence through Simulated Evolution (1966), evolved finite-state machines that predicted sequences. Ingo Rechenberg and Hans-Paul Schwefel’s evolution strategies, begun in Berlin in 1964, optimised engineering designs by mutation and selection. John Holland’s genetic algorithms, formalised in Adaptation in Natural and Artificial Systems (1975), added recombination of strings and a theory of why it works, the schema theorem. John Koza’s genetic programming (1992) evolved computer programs themselves. The umbrella term “evolutionary computation” came into use in the early 1990s, when the programmes began to share conferences and journals. Their mechanisms and theory are set out under Complex adaptive systems: methodologies.
The apparatus by design
What sets this home apart is that the whole process is designed, heredity included. The designer chooses how faithfully a solution is copied, how often and how much it mutates, whether and how solutions recombine, and the fitness function by which they are scored. Turing’s identification already put the experimenter’s judgement where the environment stands in natural selection. The process has the ingredients Lewontin’s conditions name, variation, differential success and heredity, and has them by construction; the field’s own questions are about which settings of those ingredients search well. The ingredients themselves are on the criterion page.
See also: Evolution (the subject landing) · The criterion · Complex adaptive systems · Holland · Turing