Intelligent distributed design systems: A machine-learning approach

Making the designer aware of the downstream impact of early design decisions is a critical aspect of concurrent design. Conventional "Design for X" approaches generate feedback by relying on systems with hand-crafted expert information that can be difficult and time-consuming to set up and maintain. This paradigm is ill-suited for scaling up to handle the vast space of possibilities generated by the ambiguous nature of conceptual design requirements for today's complex products and there is a pressing need for automated methods to generate synthesis information to guide conceptual design.;The solution strategies developed in this thesis demonstrate that techniques from machine-learning when framed in an appropriate design context--can be powerful and effective tools for learning information relevant to conceptual design from archives of detailed design data. Specifically, we utilize Adaptive Resonance Theory based algorithms for real-time structuring of design case information. This is applied in the context of creating flexible meta-models associating design specifications with design solutions during interaction with a design database. We further present a novel application of neural-networks utilizing reinforcement-learning to the task of learning to associate early design states with utilities indicating their downstream impact on the final design objective. The functionality of this approach is demonstrated in the context of learning over catalog-component descriptions to predict the progressive impact of specifying a component on issues like feasibility, weight etc. Finally, we describe a prototype intelligent design system--the Concept Database--developed in the course of this work as a framework for integrating our learning strategies with access to modular design case and component catalog information. The system utilizes relational database technology with a world-wide-web interface for distributed access to design models, templates and design cases. Overall, the multi-strategy approach taken in this thesis takes a significant first step in addressing the unique information needs of conceptual design by combining learning techniques from artificial intelligence with current information access technology to better leverage the advantages of making high quality decisions early in the product life-cycle.