Heuristics for Formulating System Design Optimization Models: Their Uses and Pitfalls Anand P. Deshmukh * , Deborah L. Thurston † and James T. Allison ‡ * Graduate Student, adeshmu2@illinois.edu † Professor, thurston@illinois.edu ‡ Assistant Professor, jtalliso@illinois.edu Department of Industrial and Enterprise Systems Engineering University of Illinois at Urbana-Champaign Urbana, IL 61801 Abstract—Modern systems are becoming increasingly complex due to the need to build inherent autonomous decision making abilities in them (e.g. self driving cars, fly-by-wire aircraft systems, automatically controlled wind turbines etc.) and other factors. Like many other new trends, these systems challenge the limits of existing design methods. Traditional normative design optimization methods address objective trade-off issues by employing an organized constraint matrix for quantifying the unavoidable cause and effect relationships between designer choices and product performance. Normative methods can also be used to address subjective tradeoffs under uncertainty with this same degree of analytic rigor. New design trends challenge these and other existing methods by presenting the designer more and more options and tradeoff choices that are well outside the traditional boundaries of analysis. The usual response is to expand the frame of analysis, relying on expert heuristic “rules of thumb” to make the task manageable. However, these heuristics can create unnecessary constraints or lead to cognitive biases. This paper presents a new framework for examining the steps in formulating a design optimization problem, and determining when to keep heuristics for their efficiency and when to replace them with a normative decision approach. An illustrative example employing an automatically controlled wind turbine is provided to demonstrate how a reasonable-seeming heuristic for defining of the design objective can lead to a result that is sub-optimal in terms of satisfying the true system performance objective. I. I NTRODUCTION Autonomous systems are playing an increasing role in day- to-day life and challenge the limits of existing design methods, presenting the designer with choices, tradeoffs and constraints never before considered. Advances in systems design method- ology are often made by expanding the frame of analysis to include new issues [1]. Normative design approaches consider these new tradeoffs under uncertainty with the same degree of analytic rigor that was previously applied only to static physical system modelling. While successful, there are times when the cost of normative approaches are too high. This occurs when the effort required to formulate the model, gather data to test hypotheses about the design decision makers’ willingness to make tradeoffs, assess their attitude towards risk, and formulate other elements of the analytic model, etc., are simply not worth the effort. On the other hand, the heuristic ‘rules of thumb’ employed by design experts, which is a descriptive design approach, are faster and more efficient. However, sometimes these heuristics can lead to suboptimal results because of systematic cognitive biases that lead the design choice away from the optimal solution. Heuristics also account for factors in design decisions in a simplified way, also leading to suboptimal decisions. This tension and synergy between normative and descriptive approaches to engineering design is described in [2]. The problem addressed in this paper is that a strictly heuristic design approach can lead to inferior solutions, while strictly normative approach can be too time consuming and costly to implement. This paper presents a set of rules for determining the best combination of both the approaches. II. BACKGROUND AND PROBLEM DEFINITION It is not always obvious how and when to consider emerging trends (such as design for autonomy), in the product design process. For example, [3] addressed the relatively new problem of hybrid electric vehicle battery design. They developed a method to address two criteria relevant to cell spacing design; closeness to target temperature and evenness of temperature distribution. Their approach develops the Pareto optimal fron- tier, and then employs equitable conflicting objective opti- mization to determine the best location on the Pareto optimal frontier. This computationally intensive approach identifies the optimal cell spacing, but does not yet include consideration of other decision variables such as shape of the battery pack and geometric arrangement of cells. Analytic methods have brought mathematical modeling rigor to design problem formulation steps that were previously ad-hoc. One such method focuses on determining which attributes should be considered during engineering change evaluation [4]. They formulate a multi-objective optimization problem to quantify the relative importance of feasible at- tribute sets. The goal is to identify which past design changes should be retrieved (and evaluated) to best evaluate the effect of a proposed design change. Another approach is the use of Design Structure Matrix (DSM) based methods. A DSM displays the relationships