A Novel Analysis Framework for Evaluating Predisposition of Design Solutions through the Creation of Hereditary-Amelioration Networks Derived from the Dynamics within an Evolutionary Optimizer
Michael J. Sypniewski · Deep Blue (University of Michigan) · 2019
In early-stage design, critical decisions are made within a limited information environment. Designers conduct and interpret analyses, while considering any associated risks, so that meaningful design trade-offs can be investigated. To aid in this, design tools are utilized to generate solutions in the hopes of characterizing a design space. However, while it is known that solutions are prescribed by the tools used to generate them, there seems to be little concern toward how these predisposed biases affect the quality of solutions relative to the desired outcome. Without the ability to determine a tool's biases, one cannot understand their effects on decision-making. This inability can promote inaccurate perspectives of the desired design space, can negatively impact the ultimate success of the design, and poses a currently unquantified risk within the design process. To make truly informed decisions, designers must be able to assess a tool's inherent biases, its intended applications, and its contextual appropriateness to the design questions it is being used to answer. To provide these capabilities, this thesis presents a framework for evaluating a model's underlying biases. Within this thesis, new and novel aspects of quality have been developed, namely solution-centric quality and generative quality. Novel quality metrics have been created and are used to evaluate a model as it is subjected to biases. A modified Genetic Algorithm (GA) has been developed so that an ensemble of biased solutions can be generated over time. Thereby, this GA provides a dynamic environment of the solution generation process associated with a model. Additionally, newly created and novel Hereditary-Amelioration Networks (HANs) are derived from the dynamics within this modified GA. The HANs capture the implied-causality behind solution dynamics and represent temporal, causal relationships between solutions. This newly developed approach is used to establish the comparative context necessary for a model's biases to be analyzed and understood by creating reference and biasing cases. Utilizing the developed framework, a variety of solution-centric and generative analyses are developed to evaluate a model's inherent tendencies. A comprehensive case study is conducted to demonstrate how the framework and the various analyses are implemented and interpreted. The case study's results demonstrate that the framework can successfully identify a model's biases, providing designers with the contextual information necessary to make truly informed decisions.