Join our welcoming International Design Community.
Complete this application form to join now.
Author: Potter, Stephen
Supervisor: Culley, S. & Chawdhry, P.
Institution: University of Bath
This thesis is concerned with the automation of the conceptual design of engineered artefacts and systems. During the conceptual design phase, an initial solution is generated in response to the design specification; the quality of this initial solution has considerable influence on the success of the whole enterprise. In complex domains, this task is performed by an expert who has a thorough understanding of the domain and much experience of design.
As a task requiring both skill and knowledge, conceptual design has been a focus of research by the Artificial Intelligence (AI) community, which attempts, as a more general goal, to emulate intelligent behaviour in computer systems. While several successful automatic design systems have been developed as a result of this research, a limitation common to much of this work has become apparent. On the whole, the approaches have tended to rely on knowledge engineering techniques to provide the requisite knowledge. These techniques try to capture this knowledge directly from a human expert, through structured interviews or other knowledge acquisition methods. However, knowledge engineering has been found to be lacking for the capture of heuristic knowledge, the generalised, experience-derived ‘rules of thumb’ that allow experts to go about their tasks. Unfortunately, this heuristic knowledge is essential for successful conceptual design.
What is needed, then, is some approach that is able to capture these heuristics, and in such a way as to allow their subsequent re-use within computer models of the conceptual design process. The foundations of the research described in this thesis lie in the conjecture that examples of designers’ work might prove a more profitable source of this heuristic design knowledge than do the designers themselves. It is hypothesised that a design example, consisting of a design specification and the corresponding design solution, contains implicitly the heuristics that the designer applied to effect the generation of solution from specification. This research concerns finding a way of exploiting this implicit design knowledge using existing AI techniques.
The working domain is that of fluid power systems. Conceptual design in this domain is a configuration design task, requiring the selection and connection of a set of standard domain components into a system that will meet the given specification. An archive of design examples in this domain has been constructed; so too has a means of describing these design problems and solutions in a computationally tractable fashion.
Artificial Intelligence and Conceptual Design Synthesis
The application of inductive machine learning algorithms provides a means by which the implicit heuristics might be accessed. These algorithms have arisen from AI attempts to mimic human-like learning; each algorithm attempts to acquire, in its own particular manner, a generalised description of some concept from the evidence contained within a set of examples of that concept. As such, they would seem to hold some promise for the capture of the heuristics from examples of their application. These learned heuristics could then be incorporated within a computer system providing the appropriate mechanisms for invoking and applying this knowledge.
As will be seen, algorithms of three types - artificial neural networks, classification construction and conceptual clustering - have been applied to this learning task for the design problem in hand, and a number of intelligent design systems incorporating the learned heuristics have been constructed. These systems are able to generate seemingly good design solutions in response to certain design problems; however, their knowledge contains errors and is incomplete, resulting in unsatisfactory overall performance. This poor performance might be attributed to the nature of the available design examples, but is probably better ascribed to the lack of sophistication of the algorithms themselves, which display learning capabilities far below those of humans.
In response to these problems, a second approach to exploiting the heuristic knowledge embodied in the design examples has been devised and implemented. This is termed Case- Informed Reasoning, an adaptation of the Case-Based Reasoning paradigm of problem solving. This approach involves developing solutions in poorly understood domains through analogical reasoning - solutions to new problems are suggested by analogy with the manner in which similar problems were solved successfully in the past. So, rather than attempting to learn generalised heuristics from the examples, instead the information contained within specific examples is used to inform the design decisions at the time at which they are made.
This Case-Informed Reasoning method requires a greater amount of background knowledge of the domain, but it is consistent in its ability to generate good solutions. The method would seem to present a practical way of utilising the knowledge contained in the archive given the current level of understanding of design processes, and of human intelligence in general.