The use of techniques similar to biological evolution is common in many fields. For example, consider the design of a scramjet aerospacecraft. Given the shape of the airframe it is relatively easy to compute the flow of air or plasma over it, and so determine its flight characteristics. Finding the 'best' shape for the plane is a much harder problem: the only method is to make an educated guess about an improved shape and run another simulation. To find an equation that links the vast number of variables involved in the shape of the airframe to the flight characteristics so difficult that it is effectively impossible except with the simplest of shapes. This system of guesses can be improved on by evolving designs.
In a genetic algorithm a 'genotype' is set up that determines the shape of the scramjet and hence the flow pattern. The programme starts with a design produced by a team of designers or an AI and description of the desired flight envelope. A simulation of each design is run and a 'fitness' assigned to each that depends on how close its characteristics are to the target. The best designs are used to create a second generation of hybrid genotypes (with some degree of mutation) and the process is repeated. After a number of generations the design will tend towards a locally best design (a design from which all changes result in worse performance). Running a whole series of such evolutions from radically different starting points will yield a set of such locally optimal designs, from which the best can be selected.
The future of Ad Astra