UC Santa Cruz

Enterprises that provide technical support for network engineering products and technologies are continuously accumulating terabytes of unstructured customer data in the form of service requests, device logs, bug reports, and network configurations. A large body of work in knowledge engineering (artificial intelligence, machine learning, data mining, and information retrieval) has been developed and applied within the context of important application areas such as expert systems, search engines, and more recently recommender systems. However, a relatively small body of knowledge engineering (KE) work has addressed methods and application of KE to the design and development of engineering systems and products. Furthermore, a relatively small proportion of the KE research applied to engineering systems has addressed all of the following three issues: 1) the structuring of the engineering subject-matter domains to which the theory is applied, 2) the proper integration of this domain structure with analytical KE methods, and 3) the KE software necessary to support the design and development of engineering systems.

This thesis addresses the theories, application, and implementation of knowledge engineering methods for using this collected data to improve product design, development, and delivery. To address the aforementioned three issues we have formulated a cognitive science-based representation framework for problem-solving, consisting of five sequential layers, or stages, which enables integration of diverse domains such as machine learning, product design and development, software engineering, and the statistical design of experiments.

To demonstrate and test our theories we have applied this representation-based framework, called the Integrated Meta-Representational Model (IMRM), to solve three important product design, development, and delivery problems related to computer networks. The first problem involved combining computer network domain knowledge with analytical methods from data mining and time-series analysis in order to monitor and assess the quality of a computer network security product. The second problem involved predicting whether or not an incoming customer support case should be escalated in priority in order to be resolved in a timely and cost-effective manner. For this problem we used a statistical Design of Experiments approach to optimizing the machine learning model for predicting whether or not a service request needs to be escalated. The third problem involved the development of a Knowledge Engineering Software Product for supporting the extraction of problem-solution pairs from customer service requests in order to create new computer network products and services. The work concludes by indicating how the Integrated Meta-Representational Model can be used to solve even more complex problems, involving the integration of all the core activities in engineering: design, analysis, experimentation, and prototyping/manufacturing.