Hierarchical and interactive parameter refinement for early-stage system design

Abstract

Though simulation models are extensively used for detailed design analysis, their potential benefits for supporting early-stage design have not yet been harvested. This thesis presents a new design methodology, called Hierarchical and Interactive Decision Refinement (HIDER), which enables detailed simulation models to be utilized for early-stage design of engineering systems. HIDER uses a machine-learning approach for forming fast empirical models at different levels of abstraction from the detailed simulation models. Models for different subsystems and from different perspectives are formed in a uniform representation and are used together in a single design environment to provide decision support for system design.

HIDER uses an interactive refinement approach, based on explicit use of multiple competing objectives, to enable the designer to quickly and effectively explore the overall design space. Multiple objective optimization produces sets of Pareto-optimal designs, and the tradeoffs between the different design and performance attributes in these sets are used to interactively refine a large initial design space guided by domain-independent heuristics as well as domain-dependent knowledge.

A prototype implementation of HIDER, which integrates the adaptive modeling and the interactive refinement approaches with an artificial intelligence based design environment, has been developed to demonstrate and evaluate the methodology. Results from the evaluation of HIDER for the parametric design of a diesel engine are presented. The thesis also demonstrates the methodology for the system-level design of a wheel loader simultaneously from competing perspectives, using detailed and disparate simulation models for cycle time and stability analyses.