Eecient Genetic Programming for Finding Good Generalizing Boolean Functions

This paper shows how genetic programming (GP) can help in nd-ing generalizing Boolean functions when only a small part of the function values are given. The selection pressure favours functions having as few subfunctions as possible while only using essential variables, so the resulting functions should have good generalization properties. For eeciency no S-expressions are used for representation, but a special case of directed acyclic graphs known as ordered binary decision diagrams (OBDDs), making it possible to learn the 20-multiplexer.