Tuesday, January 29, 2008

What does an artificial design agent mean by being "situated"?

Gregory J Smith and John S Gero

Key Centre of Design Computing and Cognition, University of Sydney, NSW, 2006, Australia


Schön described designing as a "conversation with materials conducted in the medium of drawing". Both the problem and solution of many designing tasks emerge through this "conversation" between a situated designer and the medium of the design. Unfortunately, describing agents as "situated" means different things to researchers from different fields. In this paper we review work from different fields so as to describe what "situated" means for a design agent.


Designing is the conscious effort to impose meaningful order (Margolin, 1986, quoting Victor Papenek). Conceptual designing is an early phase of design which is characterised by abstractness and an incomplete understanding of the problem and/or solution (Gero, 1998). Designers cope with this by exploring the space of design requirements at the same time as they begin to try and understand the space of conceptual designs. This is achieved by interacting with the media of the conceptual designs as exemplified by Schön's (Schön and Wiggins, 1992) "conversation with the medium". Designers, human and artificial, have therefore been described as situated agents. But what is meant by "situated" varies across disciplines. What Suchman (Suchman, 1987) understands by the term "situated" has a cognitive and sociogical character that is distinctly different from a common AI understanding that equates it with "embodied". So, when talking of design agents, what does "situated" mean?

The traditional computational approach presumes that designing is search and planning; achieving design goals through internal reasoning with inference rules over models in a suitable logic or language. These methods search an encoded space for a goal state, and require good heuristics to be effective. There is no notion of interaction revealing alternatives not encoded in the space. It is an approach characterised by (Coyne et al., 1990).

A logical proposition, however, is not necessarily the same as a design proposal as design problems often cannot be comprehensively stated (Lawson, 1997). Some aspects of a particular design problem do not emerge until an attempt has been made to solve it. So how does search or planning in a solution or plan space account for designing if knowledge of the design environment or design goals are incomplete or in error, or if design knowledge itself contain errors/omissions, or is incomplete, or if the design task can be formally stated but planning is exponentially complex and over a large solution space?

The heuristics that are required of any search are a part of the domain and common-sense knowledge available to a designer. Expert systems programmers have been trying to program disembodied common-sense knowledge for decades (Horgan, 2004). One reason suggested by Brooks (Brooks, 1995), Clancey (Clancey, 1997) and others for why this has been a struggle is precisely because their systems are not situated and embodied. By contrast with conventional planning, in a situated view plans are constructed as an artifact of "reasoning about action, not the generative mechanism of action" (Suchman, 1987, emphasis is Suchman's).

Fig. 1. Man standing on window ledge, from (Milligan and Shand, 1996).

An example is of a designer and sketches made during the early conceptual phases of a design task. There is a difference between this viewed as an interaction with a drawing and, say, viewing it as searching of encoded model of a drawing. One difference is that expectations of what is in a drawing influence how it is perceived it, and this influence feeds back into ongoing perceptions of that drawing. Consider the scene shown in Figure 1. We naturally believe that the man is contemplating jumping from the ledge. Now look at Figure 2, which is the same scene a few seconds later. The reason that this is funny is that it contradicts our expectations.

Consider now Figure 3(a). We do not simply look at Figure 3(a) and parse what is sensed into objects; we interact with the figure. Biasing our perception are expectations of what will be perceived. The concepts that Dali had in mind when he produced Figure 3(a) most likely include concepts of a greyhound, the mythological beast and so on as well as others that associate in his mind with those.

What we conceive of while interacting with this figure, as a viewer and independent agent, is not necessarily as Dali intended. Indeed it is not assured that what the painter conceives of afterward is only that originally intended. Many people viewing Figure 3(a) for the first time will not find all of the interpretations intended by Dali without the assistance of the Figure 3(b) sketches. Equally, just because Dali produced the image does not mean that those six interpretations are exhaustive or even necessarily correct. How we interpret the figure depends on our expectations, the current situation, and we construct the memories, beliefs and expectations that bias our perception.

Our work has therefore been motivated by a desire for a model of designing that is based on interaction; of a situated agent that can interact with an external representation of a developing design. It is the intention of this paper to describe situated design agency in the abstract, not applied necessarily to human designers or artificial agents, so as to inform the future development of artificial agents. Descriptions inspired by human behaviour are therefore intended only to that end. We do not intend these descriptions to be taken as a cognitive model of human behaviour. Computational details in this paper apply to artificial agents but will be informal (in a computational sense); a forthcoming paper will introduce formalism to this discussion. In this paper we consider research that at first blush may seem disparate. The research reviewed in this light comes from AI, computer science, cognitive science and philosophy. The common theme is a situated, interactive approach to intelligence and problem solving. We consider these ideas in the light of the actions of human and artificial agents so as to determine what it means to say that a design agent is situated.

(a) "The Endless Enigma", Salidor Dali, 1938, reproduced from (Descharnes 1985)

(b) Sketches by Dali of the images overlayed in "The Endless Enigma". They are (i) Face of the Cyclopean, Cretin (ii) Greyhound (iii) Mythological beast (iv) Philosopher reclining (v) Mandolin, compotier, figs on a table (vi) Woman seen from the back mending sail sail (Descharnes, 1985).

Fig. 3. Dali's "The Endless Enigma"

The above copied from: http://people.arch.usyd.edu.au/~john/publications/2005/05SmithGeroDS.pdf

No comments: