BREEDING INTELLIGENT AUTOMATA.

Automata populations faced with rather general computational tasks, e. g. , recognize whether or not pairs of successive inputs can be matched by amplitude scaling, are studied. The populations are subjected to tests involving execution of sample problems; each machine is given a grade and, accordingly, either gets scrapped from the population or is allowed to contribute to the next generation an almost identical automaton. A number of parameters (characterizing the connectivity of the machine) are transmitted during reproduction, but not quite faithfully. The populations learn to perform progressively more difficult tests, involving generalization. It is found that the architecture of satisfactory automata is variable and not easy to understand.