Genetic programming
Koza applied evolutionary selection not to the parameters of a solution but to programs themselves, represented as expression trees.
Why it matters
Search moved up a level: what changes is not an algorithm's settings but its structure.
Programs are crossed by exchanging subtrees and mutate by replacing a random subtree; fitness is judged by running them. The method is already described in a Stanford report of June 1990; the 1992 book gave it its settled form. It demands enormous computation and gives results that are hard to read, but does not require the form of the solution to be fixed in advance. Koza and colleagues later reported structures that infringe, improve upon or duplicate the functionality of 21 previously patented inventions, and called genetic programming an automated invention machine.