Model evolution for the realization of complex systems

Abstract

George Box said, “All models are wrong, but some are useful.” In the design of complex systems, types of complexity need to be managed. Giving the complexities that a decision maker may encounter, corresponding adjustments or improvements should be made to the design. In this dissertation, it is defined that all kinds of engineering design are comprised of four stages – formulation, approximation, exploration and evaluation – and the four stages form the model evolution loop or design evolution loop. By running the design evolution loop iteratively, a designer can handle the complexities and improve the design. Such improvements include but not limited to more robust to uncertainties, more efficient in design evolutions, easier interpretations of phenomena, etc. In the design of complex systems, as lack of data and information, heuristics are used to proceed the design, so that designers can explore the solution space and gain insight to improve the design. Those heuristics include but not limit to model structures, sub-problems identification and integration, approximation rules, and scale of details incorporated in the model. There is lacking mechanisms to evaluate the quality of the design associated with the heuristics. In this dissertation, it is hypothesized that by running the design evolution loop and exploring the solution space, designers can do the things as follows to improve the design. • Evaluating system performances associated with various heuristics (structure of the model, critical parameter setting, rules making, etc.). • Replacing the heuristics with insight obtained from exploration of the solution space to improve the design. • Managing the complexity of module structure, such as analyzing and simplifying the structure of a large number of goals. • Interpreting the behavior and the property of the model into the knowledge that supports the decision making. • Capturing and managing newly observed properties or a more detailed complexity that are not incorporated into the modeling at first – the emergent properties. • Automating the steps in the above. The intellectual merits in this dissertation are the expandable computational framework for designing complex systems and managing multiple types of uncertainty– the design evolution loop, and the methods fitting into it. By using satisficing strategy and incorporating machine learning to explore the solution space, heuristics in each of the four stages (formulation, approximation, exploration, and evaluation) can be updated or replaced by knowledge gained from experiments, calculations and analyses. In addition, knowledge on tradeoffs between different categories of design requirement – such as (but not limited to) approximation accuracy, computational complexity, design preference diversity, reformulation flexibility, and the degree of design automation – can be collected, stored and reused.