Program Induction, Complexity and Occam's Razor

ISBN-13: 9783838389349 / Angielski / Miękka / 2010 / 156 str.

Woodward, John

Opis

Search is a broad machine learning method where solutions are generated and tested. We focus on evolving computable functions with genetic programming. The literature reveals the complexity of programs is small, indicating a limitation of current methods. No Free Lunch is not valid for machine learning as simpler functions are represented more frequently which is also related to Occams razor. We argue for Occams razor, not on grounds of simplicity but probability. The complexity of a function depends on the primitives available. If the representation can build new primitives, then the complexity is independent of the primitives. We give bounds on these constants and argue these are the tightest. We examine representation, genetic operators and fitness functions. A representation which addresses a general problem is fruitful as large instances can be solved by evolving solutions to small instances. Different versions of a fitness function are compared which take into account if a program was terminated. A crossover operator is introduced which acts on modules and increases the probability of generating correctly terminating programs.

O autorze

Search is a broad machine learning method. We show No Free Lunch is not valid for much of machine learning, and conversely argue why Occam's Razor is valid. We prove results concerning the complexity of functions. Finally we examine the impact of representation, genetic operators and fitness function on genetic programming.