A study in massively parallel genetic algorithms with application to image interpretation
Description
The genetic algorithm (GA) constitutes a robust, domain-independent, and adaptive computational search mechanism. Amenability to parallelization is a major strength of the GA. The algorithmic form of the GA conforms well to SIMD computing environments with minor adjustments. This work addresses the question of the degree to which control parameters affecting subpopulation intercommunication impact the behavior of a massively parallel GA in a SIMD environment using ANOVA methods. The goal is to supplant anecdotal experience with statistical evidence. Specifically, the effects of the parameter--topology, migration operator, migration radius, and migration probability--were studied using seven response variables measuring different aspects of diversity, schemata propagation, and efficiency. This allowed insight into the behavior under various parametric conditions. Holland's revised Royal Road functions constituted the test suite. The experiment confirms much of the current wisdom but also adds details not previously considered. The empirical but rigorous conclusions provide a better knowledge of which operators and parameters to emphasize Application of the GA to scene-based image analysis was a motivation leading to the study of parallel GAs. Successful application of the GA to the task of image interpretation requires a framework that facilitates the utilization of domain-specific knowledge in computing fitness in a consistent manner over a variety of image domains. Such a domain-independent framework is developed in this work. The model uses a semantic net to capture the salient properties of the components of a typical scene from the domain. Instantiating the semantic net with the components of a candidate interpretation of the image provides a basis for estimating the fitness. The feasibility of the model is demonstrated using two distinct domains