Abstract:     4th Design Thinking Research Symposium


Abstraction in Analogy,
and Analogy in Simplification:
The Role of Representations
Extended Abstract

Design
In our research group we are concerned with Engineering Design, mostly in the Mechanical Engineering domain [AIDG 1997]. We are concerned with modeling high-level reasoning processes that occur in designing, with building knowledge-based systems that demonstrate these processes, and with experiments using these systems. The Artificial Intelligence paradigm forces us to be precise -- if you can't code it then you dont understand it enough -- and gives feedback in the form of the performance of the system.

Simplification
It's often the case that terms we use frequently and freely hide limited understanding of the knowledge and reasoning that support them. Examples of these are ``function'' [Umeda & Tomiyama 1997], ``emergence'' [Chase & Schmidt 1998], and ``simplification''.

The simplicity of a design is measured relative to some process. A design is simpler than another if a chosen measure of it is smaller. Processes include manufacturing, assembly, understanding, description, use and recycling. Measures include the number of assembly operation, the complexity of the shape, the number of components, the number and complexity of the inputs required during use, and the complexity of the description.

Simpler designs lead to fewer resources being expended during designing, manufacturing or use. Relative to the process being considered, a simpler design is a cheaper design. This makes simplification and the detection of simplicity an important aspect of designing.

However, `overall simplicity' only exists as a tradeoff between simplicity measured from different points of view. For example, simplifying a design to obtain simpler manufacturing may lead to less simple recycling.

Ultimately, simplifications are realized in the structure, function or behavior of the designed artifact itself. As we consider a ``design'' to be a potentially realizable description of an artifact, then simplifications can be recorded by altering descriptions of the artifact's structure, behavior or function.

In order to build a computational model of simplification we need to use a representation that includes and integrates all three of these `layers' [Chandrasekaran 1994].

Analogy
Simplification might be done by `noticing' some unnecessary complexity, and then applying a sequence of small transformations, guided by the goal of simplicity. More powerful, effective, and efficient simplifications might be done by analogy [Goel 1997] with previous simplifications.

These may not even be in the same domain. Provided there is an appropriate analogical match, such as between ``water'' and ``electricity'' via the shared property ``can flow'', old simplifications can be reused.

Our research investigates how simplification by analogy can be modeled. Analogy requires matching the given, target design against stored previous designs, finding a matching design that also has a record of its simplification associated with it, transferring that simplification to the target design, and adapting the resulting design to bring it back to the language and level of the target.

As simplifications of structure, behavior or function are possible, analogical simplification can work at each of these levels. However, once a simplification (e.g., behavioral) has been completed, the consequences need to be propagated to the other levels (e.g., functional and structural). This can be guided by the descriptions at those levels in the target design, the descriptions at those levels in the simplification of the matched design, and by other general knowledge that relates typical descriptions at one level with typical realizations in another.

Abstraction
Abstraction plays several key roles in the simplification process.

It is required in order to produce a description of the given design that is at a high enough level of abstraction that cross-domain matching can be easily accomplished. This will need to be guided by the goal of simplification at a particular layer (e.g., structure), for a particular process (e.g., assembly), with a particular measure (e.g., operations).

A trace of this abstraction can be used, in reverse, to guide adaptation of the design that results from applying the simplification found.

In addition, abstraction of the simplification that is found may be necessary before it can be applied to the given design. It is hoped that eventually a collection of abstracted simplifications can be generalized to form `simplification principles'.

Representation
As simplifications can be made by altering descriptions of the artifact's structure, behavior or function we use a three layer representation for each design.

The description of Function -- which we take to be the designed artifact's intended interaction with an environment, possibly a user -- is supported mainly by the Behavior description (e.g., state changes detectable outside the artifact), but possibly also by properties described in the Structure layer (e.g., color, reflectivity, or sharpness).

Function is currently described by inputs and outputs, their type and complexity, and their sequence. Behavior is described in terms of state changes or causal chains. Structure refers to components, and their properties, and the relationships between components.

The representation must support abstraction in order for the all the analogical simplification processes to take place. The hierarchies of designs that result -- and there might be many dynamically formed hierarchies -- mediate between the designs from different domains.


References
Artificial Intelligence in Design Group, web page, http://www.cs.wpi.edu/Research/aidg/, WPI, Worcester, MA, USA, 1997.

B. Chandrasekaran, Functional Representation: A Brief Historical Perspective, Applied Artificial Intelligence, Special Issue on Functional Reasoning, Vol. 8, No. 2, 1994, pp. 173-198.

S. Chase & L. Schmidt, AID98 Workshop on Emergence in Design, web page, http://www.arch.usyd.edu.au/~scott/AID98/emergence-workshop/, 1998.

A. Goel, Design, Analogy and Creativity, IEEE Expert, Vol. 12, No. 3, May/June 1997, pp. 62-70.

Y. Umeda & T. Tomiyama, Functional Reasoning in Design, IEEE Expert, Vol. 12, No. 2, March/April 1997.


Biographical Information:
DCB
MEB


Version: Wednesday July 29 20:11:35 BST 1998