Model-based support for mutable, parametric, system-level design optimization

The design of dynamic computer controlled electromechanical systems with multiple and often conflicting optimization objectives presents formidable challenges. In many real world design problems, the relations between design variables and performance parameters are mutable, i.e., they vary with input tasks and system variables. Traditional methods for parametric design optimization in Operations Research and Artificial Intelligence, which assume that the relations between performance criteria and design variables are expressed in the form of invariant algebraic functions, do not apply to mutable optimization problems. Developing globally optimal design solutions requires a system level view of the design problem which accounts for dynamic interactions among the system's components in analyzing and optimizing its behavior for different system workloads.;This dissertation develops an intelligent support system to assist human designers in solving system-level, mutable, parametric design optimization problems. Given a description of the system's components and its configuration, we determine values for its design parameters that optimize specified objectives while meeting specified design constraints. We address the issue of system-level parametric design optimization by developing the framework for a compositional component-oriented system representation which facilitates reasoning about its behavior in a holistic manner. We develop model-based reasoning techniques to address two primary tasks that are key to mutable design optimization. The first task employs a structural model of the system to dynamically generate mutable optimization relations between the system's optimization objectives and its design variables. The second task employs sensitivity analysis techniques on the derived relations to efficiently navigate the design space in search of a good solution. We perform empirical analyses to demonstrate the effectiveness of our design optimization methodology using examples from the domain of reprographic machines.;This dissertation formally characterizes the mutable design optimization problem. The Environment Relationship net framework is adapted to build discrete event models of system behavior at suitable levels of abstraction that can be used for the purpose of system-level mutable parametric design optimization. Empirical analysis demonstrates that our design optimization methodology is both effective at finding good solutions and it does so quite efficiently by pruning large portions of the design space. This is achieved by the use model-based reasoning to guide the search process to promising regions of the design space without exhaustive search, the employment of cluster analysis to reduce the number of jobs considered during the optimization process, and the randomization of the algorithm that searches for optimal solutions to eliminate local optima.