| MultiResolution Analysis (MRA) is a common phenomenon of human intelligence. The basic procedure of MRA is that a series of analyses are carried out on an object's representations at different; progressively increasing resolution levels and analysis results at lower resolution levels act as guidance to analyses at higher resolution levels. Many methods such as coarse-fine template matching, reduced model in optimization, coarse-fine path-finding and so on can be seen as implementations of principles of MRA. This dissertation reports on investigations of MRA in different areas from a unified perspective and proposes algorithms from the viewpoint of MRA for attacking machine intelligence tasks in areas such as classification, function approximation, rule learning, optimization by stochastic search, control, and so on.; The focus of this dissertation is about three questions: firstly, what is multiresolution? secondly, how to obtain multiresolution representations of an object? and thirdly, how to utilize results attained at low resolution levels as guidance for analyses at high resolution levels? Two new concepts, resolution and scale of a cluster, are introduced in this dissertation, and based on these concepts clustering algorithms are developed for obtaining multiresolution representations of an object. Using this uniform approach to attaining multiresolution representations, implementations of MRA are discussed and illustrated with examples for various machine intelligence tasks. |
