MARCH-APRIL 1997 (Vol. 12, No. 2) pp. 14-16
0885-9000/97/$31.00 © 1997 IEEE
Published by the IEEE Computer Society
Published by the IEEE Computer Society
Guest Editors' Introduction: Understanding the Nature of Design
|The problem of design|
|The special issue|
PDFs Require Adobe Acrobat
This special double issue of IEEE Expert surveys the state of the art of research in AI in design, focusing on a variety of topics that have significantly increased our understanding of the nature of design. These topics have led to the development of systems that can tackle complex design problems, as well as to new, useful AI theories and techniques.
AI in design includes the modeling of designer activity, the representation of designer knowledge, and the construction of either systems that produce designs or systems that assist designers. 1 Through these activities, we hope to gain better insight both into the nature of design processes and representations, and into methods for developing systems to support design activities. More generally, these studies contribute to our understanding of intelligent behavior.
This area has various names, including AI in Design (AID), Knowledge-Based Design Systems (KBDS), Intelligent CAD (IntCAD or ICAD), and Knowledge Integrated CAD (KIC). 2 Areas of related work include concurrent engineering, simultaneous engineering, and concurrent design.
Design researchers use a variety of AI techniques, such as constraint satisfaction, search, negotiation, and knowledge representation. Because design is inherently multidisciplinary, researchers often draw on results from fields such as cognitive psychology, decision theory, optimization, language theory, and architecture. 3, 4 The task of design (as opposed to, say, diagnosis) and the domains in which design is being done (for example, computer design and bridge design) influence how these techniques are used.
This task/domain context reveals new uses for existing techniques and provides a catalyst for the invention of new ones. For example, in a case-based-reasoning approach to design, the adaptation of a retrieved design might be more complex than in most other applications of CBR and might require new methods, because of the many dependencies between the design decisions that make up the design.
The problem of design
Design is a peculiarly human activity: the desire to recast the world to suit our purposes is a defining characteristic of being human. It is also a defining characteristic of being intelligent. 5 The study of design, therefore, is naturally part of the study of AI.
Defining design, in its broadest sense, is a difficult task. Many technical and philosophical tracts have been written on the topic; we make no attempt to wade into those choppy waters!
We loosely interpret design to be an information-processing activity that creates a description of an engineered artifact (for example, a building or a software module), guided by some set of specifications and some set of constraints. The specifications are intended to describe what is desired; they are not necessarily methods that describe how the artifact should be designed. They might vary from concrete to very abstract. Specifications can include such things as aesthetics, resources, economics of manufacturing, and performance. Constraints describe the limitations of the natural world (for example, Newton's laws) and the other, perhaps artificial, limitations to be placed on the design. They describe what must not be violated.
A design process is complex. One of the most well-studied design problems, configuration design, has been shown to be exponential in size and time. The configuration-design process selects parts from a catalog or set of catalogs to realize a set of specifications and satisfy constraints about how parts can be combined. Because configuration design is one of the simpler forms of design, more complex types of design will be likely to also have exponential complexity. Thus, with apparently intractable search, design problems are hard to solve.
Design is hard for several other reasons:
The goal state (that is, a good design) is often hard or impossible to define. Often, it is the result of the design process that identifies candidate goal states. For example, in designing a house, it is possible to initially describe desirable characteristics, but it is through the design process that possible houses emerge.
The implications of various design actions are not generally predictable. For example, if a part is added to a design, it is not possible to fully predict that part's impact on the design. The part might combine well with other parts, might lead to vastly inferior solutions, or might not interact with other parts at all.
Thus, to solve these problems, AID researchers need to draw on a wide range of AI techniques, such as knowledge representation, problem-solving methods, and machine learning. These techniques let researchers cast design problems in computable forms, allowing the underlying structure to become apparent. Through such elucidation, it is often possible to develop heuristics that make it possible to solve interesting design problems, or to find ways to bound design problems so that they can be solved.
AI methods also provide researchers with a (sometimes formal) framework for analyzing how humans make design problems tractable. For example, the natural organization of parts catalogs, with each part completely distinguishable from the other parts, can help form effective heuristics for making some classes of configuration-design problems tractable. 6, 7
The complexity of the design process demands a richness of reasoning that contrasts with the highly simplified worlds and restricted domains often used to develop and demonstrate AI techniques. Developing a knowledge base that is rich enough to allow valid experimental analysis of a design system or algorithm often requires substantial effort. Recently, researchers have created ontological descriptions of design problems. 8
In addition, complexity often makes it necessary to combine AI techniques such as multiagent systems and machine learning. 9 Researchers hope that by integrating a variety of complementary techniques, they can develop effective heuristics for realistic, complex design problems.
The special issue
This special double issue consists of a set of invited articles from AID experts. We begin with an interview with Clive Dym. Through his editorship of one of the prominent journals in the area, AIEDAM (Artificial Intelligence in Engineering Design, Analysis, and Manufacturing), he has gained unique insight about AID. 10
Complementing Dym's perspective is that of Dan Siewiorek, Steven Fenves, and Georgette Demes, of the Engineering Design Research Center at Carnegie Mellon University. 11 The EDRC has a long history of AID research, with particular emphasis on multidisciplinary research.
Bob Wielinga and Guus Schreiber describe configuration-design problems. Although configuration has been well studied, recent work has better characterized these problems and has provided greater understanding of how to solve them. Central to configuration design, as well as other types of design, is constraint solving. Several articles in this track describe approaches to resolving conflicts between constraints.
Making design systems creative, allowing them to develop novel solutions, has long been of interest to design researchers. Ashok Goel discusses how analogy is a key ingredient of creativity in design. As a complement to that article, Mary Lou Maher and Andrés Gómez de Silva Garza describe how CBR can help create designs where the design knowledge and process cannot be formally described. A challenge for CBR applications in design is finding a good indexing scheme.
Closely related to CBR, which uses remembered past designs or design processes, is machine learning. As we noted earlier, developing design systems requires substantial knowledge bases and efficient mechanisms for obtaining, using, and maintaining this knowledge. ML is vital for tackling these issues, and Alex Duffy provides an overview of ML techniques used by design researchers.
Many design researchers believe that the description of function is central to understanding design. In some design processes, especially the more creative ones, reasoning about function is primary. In configuration design, however, the notion of function is implicit. Yasushi Umeda and Tetsuo Tomiyama describe current research on the representation of, and reasoning about, function. Such research is related to analogy and CBR in that it tries to make the types of design problems that can be automated more general and less focused on detailed, parametric design.
Reasoning about function is one aspect of reasoning qualitatively. Yumi Iwasaki, in her "Expert Opinion" column, comments on qualitative reasoning's role in design. This research area encompasses the representation and use of a variety of design-related concepts, including function, behavior, causality, structure, flows, and shape.
For design researchers who work mainly in the 3D world, such as mechanical engineers and architects, structure is of central concern. Ken Brown discusses how grammars can represent structure and generate new structures.
A design rationale comprises the reasons underlying a particular design decision. Design rationale is very important when attempting to evaluate and change existing designs. Jintae Lee describes the use of rationales as a design decision-making tool.
As design researchers begin to design large systems and use human organizations (such as design teams) as their inspiration, their focus changes from an individual system to multiagent systems, where each agent has expertise for solving a part of a larger design problem. Susan Lander describes the organization of multiagent design systems, and methods used to both coordinate their actions and resolve their conflicts.
Although this special double issue does not include articles specifically about any current applications of AID, it is still of great relevance to anyone interested in the applications of AI. It is of particular interest to anyone who is building applications in this area, anyone wishing to start work in some branch of AI in design, and anyone already working in one of the topic areas. Each article provides references to various research topics, systems, and related literature.
David C. Brown is a professor of computer science at Worcester Polytechnic Institute. His research interests include computational models of engineering design, and the applications of AI to engineering and manufacturing. He has been a consultant to Digital Equipment Corporation on the application of AI to engineering and manufacturing and is on the editorial boards of AI in Engineering, Design, Analysis and Manufacturing and Concurrent Engineering: Research and Application. He is a member of the advisory committees for the AI in Design and the AI in Engineering conferences and for the IFIP WG 5.2 workshops and conferences. He is the coauthor (with B. Chandrasekaran) of Design Problem Solving: Knowledge Structures and Control Strategies (Pitman Publishing) and a coeditor of Intelligent Computer Aided Design (Elsevier Science Publishers). He has a BSc, an MSc, an MS, and a PhD in computer science. He is a member of the ACM, the IEEE Computer Society, and the AAAI. Contact him at the AI Research Group, Computer Science Dept., Worcester Polytechnic Inst., Worcester, MA 01609; firstname.lastname@example.org; http://www.wpi.edu/~dcb/.
William P. Birmingham is an associate professor in the Electrical Engineering and Computer Science Department at the University of Michigan, Ann Arbor. He is also a member of the University's Collaboratory for Research on Electronic Work. His research interests concern large, distributed information systems, including distributed optimization and design, concurrent engineering, and digital libraries. He is the lead system architect on the University of Michigan's Digital Library Project, funded by an NSF, ARPA, and NASA joint initiative. Common threads in these areas are the coordination mechanisms for a large number of agents, interoperability, and the creation and maintenance of large knowledge bases. Birmingham is an NSF Presidential Young Investigator. He is the new editor of AIEDAM and is on the editorial board of IEEE Expert. He received his PhD in electrical engineering from Carnegie Mellon in 1988. Contact him at the Electrical Engineering and Computer Science Dept., Univ. of Michigan, Ann Arbor, MI 48109; email@example.com.