A central problem in computational learning is to determine which types of Boolean functions (concepts) can be learned by efficient algorithms. The difficulty of learning can vary in different learning models. The models represent different conditions under which learning might take place. A major goal of this project is to distinguish, in a variety of learning models, between classes of Boolean functions that are efficiently learnable and classes that are not.
Among the learning models that will be studied are query models and attribute-efficient models. In query models, the learning algorithm can obtain information about the function being learned by asking queries. Query models represent situations in which the learner has access to a teacher or expert, or can perform experiments and observe outcome. In attribute-efficient models, the learner is expected to learn efficiently even in the presence of a lot of irrelevant information. The presence of irrelevant information is a common complication in applied machine learning problems such as automatic text categorization.
Another major aspect of this project concerns characterizations of classes of Boolean functions. The project aims to discover connections between characterizations of Boolean function classes and the complexity of learning and recognizing functions in those classes. This approach is inspired in part by work in graph theory. The characterization of Boolean function classes has not been studied as extensively as the characterization of graph classes. A goal of the project is to do for classes of Boolean functions what has been done for classes of graphs: to establish connections between characterizations of classes and the complexity of problems involving the classes. A focus of this project will be on classes of Boolean functions that are closed under basic operations, such as permutation of variables, and substitution of constants for variables. The motivation for restricting attention to such function classes is to obtain elegant results in learning and complexity, results that may not hold for all Boolean function classes.
