Font Size: a A A

Decentralized computing and machine learning applications in multidisciplinary analysis

Posted on:2002-07-12Degree:Ph.DType:Thesis
University:Rensselaer Polytechnic InstituteCandidate:Kim, Byong SunFull Text:PDF
GTID:2468390014450342Subject:Engineering
Abstract/Summary:
The proposed research is primarily in the areas of developing efficient approximation techniques that will reduce the high cost of computation. The fundamental approach can be classifed as belonging to the area of computational intelligence. The first approach examines the use of genetic algorithm based machine learning method referred to as classifier systems to improve function approximation. A global function approximation method such as artificial neural network entails the problem of requiring too many sets of training data to produce reliable approximation. Classifer systems determine optimal numbers and distributions of the training data. A second application of machine learning methods is in the realization of completely new methods of approximate analysis better suited for implementation in parallel machines. These methods belong to the category of cellular automata computational model that is intrinsically parallel and ideal for adaptation on parallel computers. Cellular automata are fully discrete models of a physical system. The state of each cell in the system is determined by interaction between the immediate neighboring cells dictated by local rules of interaction. Results from the work in this area indicate that discovering the appropriate local rules of interaction is the key issue. Two different ways of discovering local rules of interactions are proposed in the current study. The first approach is to use genetic algorithm based machine learning. The implementation of this approach indicates that the rules tend to be problem-dependent, underscoring the need for deriving more general rules that can be useful if cellular automata are to be used as an alternative efficient reanalysis technique in structural synthesis. The use of classifier system based learning is also adapted for rule discovery that proceeds without any pre-determined rule templates. This results in the rules that are slightly more general than those discovered through the use of conventional GA. The second approach derives the rules by taking advantage of physical principles that govern elastic solids. In structural analysis, such rules could be obtained through modeling of the domain as point masses interconnected with elastic springs. The rules obtained from the physics of solids are shown to be sufficiently general to allow for use as a reanalysis tool, and hence can be further extended to perform simultaneous analysis and design. (Abstract shortened by UMI.)...
Keywords/Search Tags:Machine learning, Rules, Approximation
Related items