Methods for Statistical Inference : Extending the Evolutionary Computation Paradigm

Methods for Statistical Inference: Extending the Evolutionary Computation Paradigm A dissertation presented to the Faculty of the Graduate School of Arts and Sciences of Brandeis University, Waltham, Massachusetts by Hugues Juill e In many instances, Evolutionary Computation (EC) techniques have demonstrated their ability to tackle ill-structured and poorly understood problems against which traditional Arti cial Intelligence (AI) search algorithms fail. The principle of operation behind EC techniques can be described as a statistical inference process which implements a sampling-based strategy to gather information about the state space, and then exploits this knowledge for controlling search. However, this statistical inference process is supported by a rigid structure that is an integral part of an EC technique. For instance, schemas seem to be the basic components that form this structure in the case of Genetic Algorithms (GAs). Therefore, it is important that the encoding of a problem in an EC framework exhibits some regularities that correlate with this underlying structure. Failure to nd an appropriate representation prevents the evolutionary algorithm from making accurate decisions. This dissertation introduces new methods that exploit the same principles of operation as those embedded in EC techniques and provide more exibility for the choice of the structure supporting the statistical inference process. The purpose of those methods is to generalize the EC paradigm, thereby expanding its domain of applications to new classes of problems. Two techniques implementing those methods are described in this work. The rst one, named SAGE, extends the sampling-based strategy underlying evolutionary algorithms to perform search in trees and directed acyclic graphs. The second technique considers coevolutionary learning, a paradigm which involves the embedding of