Μεθοδολογίες σχεδιασμού βασιζόμενες σε τεχνικές υπολογιστικής νοημοσύνης και σε προηγούμενη σχεδιαστική γνώση

Abstract

The current Ph.D. thesis introduces a methodogical and systematic approach in order to perform engineering design by utilizing methods and techniques of computational intelligence as well as past design knowledge. Important issues and processes, such as representation and manipulation of design knowledge, extraction of optimal design solution, learning and reuse of design knowledge etc. are analytically discussed. Each one of the aforementioned elements is considered as core around which integrated algorithms are developed that combine various computational intelligence (soft-computing) techniques such as fuzzy logic, artificial neural networks and genetic algorithms. The parametric design problem is formulated on a mathematical basis, whereas the collaborative formation of fuzzy design structure matrices (DSM) is introduced as a basis on which the available design knowledge is organized in convenient hierarchical structures. Furthermore, in order to model the design objectives/requirements/constraints, a fuzzy inference framework is introduced that facilitates the expression of fuzzy preferences on various design parameters. The metric of the total maximum fuzzy preference is introduced as optimality criterion for the design solutions and then it is integrated in the proposed design knowledge organizational structure. The optimal solution is extracted either through an optimization process basically using a genetic algorithm or -alternatively- other optimization techniques, or through deploying analogical reasoning design with retrieval of past design solutions. The design solutions are preserved in a case base and they are retrieved by a trained competitive artificial neural network, which classifies the solutions into clusters and extracts the cluster that converges to the current designer’s preferences. Multiple options are provided for the extraction of optimal solutions. These options include: a) an optimization process that utilizes genetic algorithms, b) other evolutionary optimization techniques and c) analogical reasoning with retrieval of past design solutions. The design solutions are preserved in a case (solution) base and they are retrieved by a trained competitive artificial neural network that classifies the solutions into clusters and extracts the cluster that suits best to the current designer’s preferences. Architectures are developed and introduced, based on which the optimization and case-based retrieval processes are combined. This combination provides more efficient results in terms of quality and speed, if a comparison is made versus the implementation of individual (non-hybrid) techniques. Additionally, a methodology is proposed, according to which the design solutions located in the case base can be retrieved and utilized for a simplified neuro-fuzzy approximation of the initial design problem. It is proved that this approximation is suitable in case of design problems with high computational cost. In the context of a systemic approach of implementing the proposed methodologies and architectures and their evaluation, a system named Case-DeSC (Case-based design with Soft-Computing) is developed and tested against overall performance in three different design problems. The results from the implementation of the proposed methodologies, their future enhancements and evolution, as well as a general discussion about engineering design conclude the present dissertation.