A Novel Framework Utilizing Bayesian Networks Structured as Logical Syllogisms to Determine Sufficiency of Early Stage Ship Design Knowledge Queries

Abstract

Modern engineering design is a complex, temporal, path dependent process in which designers generate knowledge for decision-making. Designers do this by utilizing design tools such as optimization or design synthesis tools that generate data to explore a potential design space. There is a misguided equivalence within the naval design community that data, information, and knowledge are interchangeable. It is assumed that design data is a representation of relationships within and between data, information, and knowledge. Knowledge, which is ultimately the goal, is derived from the proper combination of information with analysis and experience. Decision outcomes made based on designer perception of knowledge will not be determined until the impacts of the decision are realized at a later date.

The critical question addressed in this thesis is “how does one determine the sufficiency or quality of a design decision based upon design tool generated data without awareness of future outcomes?” To address this issue, the first realization is that all design tools are created based upon knowledge artifacts. These artifacts and their structure predicate the available data that will be used for design decision-making. Often these tools are modified or extended for use beyond their original development intent. While design data generating tools are assumed to be generalized for a wide range of appropriate use, the ability to determine if this is true does not exist within the literature. One can clearly prove failures in design tool suitability through traditional convergence statistics and lack of Pareto optimality. The issue lies in the regions in which solution generation and solution statistics, as well as Pareto front generation, is deemed successful. However, statistical success does not relate to data suitability for the desired design knowledge generation and query.

Knowledge queries are uniquely human in that, unlike computers, designers have the ability to follow nonlinear patterns of associations to generate inferences based on data perception. This puts designers at risk for formulating inferences that may not be true relative to the underlying statistics associated with how the inference data was generated. Cognitive biases define the myriad methods in which designers associate meanings to data in ways that do not directly map the data. While the thesis contained within does not directly address cognitive biases it provides a mechanism for the evaluation of design inferences utilizing logical syllogism evaluations of design space exploration data as the means for the determination of sufficiency of generated data used for decision-making.

This dissertation presents the development of a novel logic-based syllogistic Bayesian framework that enables the evaluation of the suitable use of a tool for knowledge development or queries. This work utilizes AAA1, AAA2, and AAA3 syllogisms to represent designer decisions. Syllogisms are transformed into Bayesian networks used to calculate metrics for evaluation of the data generated by an instance of a tool. A unique approach to network construction allows for Bayesian probability and mutual information to examine designer decision-making as syllogisms. Within this work, several cases will be presented that demonstrate the methodology as well as extend the framework to encompass non-standard layered logical structures.