T55.4 CSWEY .W2 no. A Compararive Analysis of Design Rationale Representations Jintae Lee Kum-Yew March 1992 Lai WP # 84-92 INTERNATIONAL CENTER FOR RESEARCH ON MANAGEMENT OF TECHNOLOGY Massachusetts Institute of Technology Sloan School of Management Cambridge, Massachusetts The International Center for Research on the Management of Technology A Compararive Analysis of Design Rationale Representations Jintae Lee Kum-Yew Lai WP March 1992 Sloan WP# 3295 CCS TR# A revised and condensed version Human-Computer © Interaction & # 84-92 3405 121 of this report appears in the special issue of on design rationale, v.6(3-4), pp.251-280. 1992 Massachusetts Institute of Technology Sloan School of Management Massachusetts Institute of Technology 38 Memorial Drive, E56-390 02139-4307 Cambridge, MA d/l.l.T. TF^ l,f'''^A»''^"1 ? 3 1993 A and condensed version of revised Human-Computer this report appears in the special issue of on design rationale, v.6(3-4), pp. 251-280 Interaction A Comparative Analysis of Design Rationale Representations Jintae Lee and Kum-Yew Lai Center for Coordination Science and MIT Artificial Intelligence Laboratory ABSTRACT A few representations have been used what ways they we for capturing are adequate or limited so that develop a framework generic design tasks. We design rationale. we know how to important It is improve them. for evaluating design rationale representations to know in In this paper, based on a set of build the framework by progressively differentiating the elements of when made explicit, support an increasing number of the design tasks. we evaluate the expressiveness of the existing representations. We also present a language, DRL, that we believe is the most expressive of the existing representations without being too complex for human users. We also discuss the limitations of DRL as open design rationale With this problems that, framework, for further research. 1. INTRODUCTION As the articles in this issue point out, an explicit representation of design rationale can bring many benefits. Such a representation can lead [McCall 1987; Yakemovic principles underlying & human-computer provide a basis for learning, The 1990a]. largely extent to were represented from what we understanding of the issues involved justification, interaction [Carroll & benefits, 1989]. we already get from the notes on paper that obtain from we It of the can also et al. 1989; Lee however, would depend for representing design rationale. in free text, the benefits and et al. 1989], Campbell and computational support [Fischer which we can actually reap these on the language we use rationale to a better Conklin 1990], of the design space [MacLean If, it for example, design would not be take in design meetings. different Also, the Jintae Lee and Kum-Yew Lai, Center for Coordination Science, 02139. Massachusetts Institute of Technology, E40-140, 1 Amherst Street, Cambridge, Electronic mail addresses: iin@ai.mit.edu and kumyew@eagle.mit.edu. Authors' present address: MA Comparative Analysis of Design Rationale Representations CONTENTS 1. INTRODUCTION 2. WHAT DO WE WANT TO DO WITH 3. WHAT'S IN DESIGN RATIONALE? 4. EXISTING REPRESENTATIONS 1 3 5 15 4.1. IBIS (Issue Based Information System) 16 4.2. Toulmin's Model of Argumentation 21 4.3. QOC 24 4.4. Other Representations (Question, Option, and Criteria) 26 DRL (DECISION REPRESENTATION LANGUAGE) 5. 5.1. Introduction 5.2. Description of 29 29 DRL DRL 30 as a Design Rationale 5.3. Evaluation of 5.4. Relation to Other Studies Language 35 40 CONCLUSIONS 6. 42 kinds of computational support that explicit to DESIGN RATIONALE? we can provide depends on what a representation makes and how formal the representation is. A few systems have been built and actually used & Rlttel 1970; McCall 1987; Conklin & Begeman capture design rationale or arguments [Kunz 1988; Fischer et al. 1989; Lee 1990a, 1990b; McCall 1990], based on the earlier studies of design activity [Kunz 1958]. However, there is no systematic attempt and most of them used representations & Rittel 1970] or of argumentation [Toulmin to justify the choice of these representations or discuss the rationale for using them. This paper is motivated by the following questions: representations? how do we Do they allow us to easily represent what How adequate are the existing we want to represent? evaluate a language for representing design rationale? This paper answer these questions by identifying the elements of design rationale explicit and by exploring the consequences of making them explicit. is In general, an attempt that could be to made Laying out these elements provides a framework for placing the existing meanings of design rationale in perspective, as well as providing a framework for evaluating a representation language for design rationale, as we hope to show in this paper. Comparative Analysis of Design Rationale Representations We proceed in the following way. we First, reference against which we we would choice of that artifact. and making resulting which explicit We what model allows us to define the framework, we artifact is associated support In Section 3, discuss. We start with a with a body of reasons for the then elaborate this simple model by incrementally differentiating implicit in the is to do. body of reasons. As we do scope of a representation and adequacy within its evaluate the existing representations in Section we representations and overcomes many we current limitations, which so, we what each discuss These models of design rationale provide a framework language, called DRL, which 2. we evaluate the representations that to will use these tasks as a characterize design rationale by presenting progressively richer models. simple model of design rationale, where an its we might want identify the tasks that Throughout the paper, we using a design rationale representation. 3 believe is its more expressive than most of their limitations. we In Section 5, 4. Using scope.^ in this present a of the existing As we describe DRL, we also discuss present as open problems for future research. WHAT DO WE WANT TO DO WITH DESIGN RATIONALE? To evaluate a representation, we need know what to tasks designed it is to For example, [Mostow 1985] different levels of abstractions. ^ lists The support. that a design rationale representation can or should support can be described in tasks many ways at the following tasks: documentation, understanding, debugging, verification, analysis, explanation, modification, and automation. [Fischer et al. 1991] points out that maintenance and redesign of an during design process. [MacLean documenting design rationale can support reuse of the design knowledge, and artifact, et al. 1991] list two major benefits The representation: aid to reasoning and aid to communication. critical reflection from design rationale tasks of achieving these benefits are elaborated further in terms of subtasks such as enabling the designers to envisage the available design alternatives in a more structured way, with the arguments for and against them. Another way of characterizing the tasks is to list accomplish the general tasks mentioned above want our design In this paper, the questions that in the rationale representation to help a we To the extent that use the terms, model and representation, in the following way. model can have several representations. structure independent of a particular we often need to answer to way use the term representation to refer we answer these questions, answering these algebraic concept, and representation a linguistic one. Since the many ways, we design process. of describing it, same Therefore, we when we want use the term model. to a particular notation for Model is an structure can be described in On to discuss a the other hand, describing the structure. There are many good discussions of the benefits or tasks for which a design ratioanle can be used. Those not discussed in this paper include [McCall 1987 ;Yakemovic & Conklinl990 ]. Comparative Analysis of Design Rationale Representations questions become the tasks that the representation should support. we representative questions that 1991a], is a set of et al. 1991], and from creating scenarios [Carroll 1990]. • What is • What did • What are the alternative designs and • What are the • The following gathered from our experiments with design rationale [Lee from walking through examples [Lewis & Rosson 4 the status of the current design? we discuss last week and what do we need two most favorable Sun Microsystems what are to do today? their pros and cons? alternatives so far? X/NeWs just released their server. How would the release change our evaluations? we do • What • Why is • What are the issues that • What are the unresolved issues? • What's the consequences of doing away with • How if not consider portability? portability important anyway? depend on this issue? What did other people deal with are we currently doing about them? this part? problem? What can this we learn from the past design cases? The above we list of the questions did not walk through and is by no means complete as there are many possible paths that many scenarios that we did not construct. We also left out those questions which, though important, do not seem to be the job of design rationale to answer How can we compute the total cost of this design?) (e.g. Nevertheless, the questions in the list provide a useful framework for assessing the expressiveness of the different representations. When we discuss the limited or increased expressiveness of a given representation, those questions that can or cannot be answered as a result. question which is not represented in the list, then we If We want to would have emphasize further that we representations of design rationale. That not mean questions. that any representation for 1). made if not, it is desirable to answer the questions in the list does design rationale should support the answering of these Each representation must weigh the costs and benefits involved These tradeoffs should what additional explicit. are assessing only the expressiveness of the existing three general dimensions: expressiveness, (Figure be to refer to can always evaluate the representations by asking whether they would support answering the question and objects, attributes, or relations we our task includes answering a in turn human usability, in tradeoffs and computational among tractability be motivated by the tasks that are intended to be Comparative Analysis of accomplished using the representation. Design Rationale Representations we In short, this paper would make the architects of a representation what their what any are not dictating we do hope representation should or should not have. However, for 5 existing that the analysis presented in design rationale more conscious of language can or cannot express and why. Domain Figure To be as sure, human that is in computer supported we cannot separate our concern with expressiveness not we Elements to capture design rationales, but much point in evaluating falls into that discuss if it is is be difficult to evaluate tradeoffs & dimensions such as expressiveness is complex from other concerns such a language for people Whether any 1991]. is to meant to be manage, then of the representations All the languages discussed Nevertheless, will we all succeed at believe it would the three dimensions without calibrating individual [Brachman (cf. if no guarantee that they Yakemovic among entirely an empirical question. here have been actually used by people, but that their "industrial strength" use [Conklin too expressiveness. its category activities. For example, usability or computational tractability. used by people there 1. & Levesque 1984] on tradeoffs between expressiveness and computational tractability for general knowledge representation). 3. WHAT'S IN DESIGN RATIONALE? What is meanings design rationale? in [Yakemovic Rosson its & 1990], Design rationale for an artifact has at least three different current usage: a historical record of the reasons for the choice of an artifact Conklin 1990], a and set of psychological a description of the design space & claims embodied by an artifact [Carroll [MacLean & et al. 1989].'^ Conklinl990] describes logical as well as historical aspects For example, the support relation has both the logical We associate [Yakemovic & Conklin 1990 ], as well as [Lee aspect and the historical aspect because record because one of their goal is to capture and with historical Potts & Bruns the 1990a; 1988] The representation used of design rationale, as in we [Yakemovic discuss further , document the actual process of design. later. Comparative Analysis Design rationale often means the Design Rationale Representations of historical record of the analysis that led to the choice of the particular artifact or the feature in question. To command like: window, specific to a A design window. menubar in the global SAVE is to put the window commands was that it saves screen space window in anyway because people such as more designed make to executing a (e.g., in the sense of [Carroll artifact is to & command. we & Rosson 1990]. and formulated like: is We can have the global menu provide more structure Such structure the set of psychological claims i.e. "Dimming commands." meaning if in a testable form. "The global learn about the "the claims that usually is embodied by an would have of design rationale they were, they to artifact be true if historical record; the claims would have the how one used by [MacLean are these alternatives related, Macintosh example, the design rationale logically possible alternatives for placing the tradeoffs are. It is need be extracted from the environment easier to use because the irrelevant items in the global is to For example, the design rationale in this sense menu makes what are in this menu makes et al. 1989], located in the space of possible design alternatives: alternatives, to each from an argument supports a proposal) and/or These claims are different from the clutter." or * decided travel a proposal replaces another proposal). Kellogg 1989], reduces screen is mouse be successful" or "the claims about the psychological consequences for the user" would be something third commands on requires a long we 20. bar at the top of the screen discuss in the rest of the paper. not be present in the historical record; even the record But, would be something by Mark on January to putting the it a generally agreed that that the advantage, together with others Another meaning of design rationale artifact opposed explicit the logical structure (e.g., the historical structure The as window the that saves the contents of the raised implementation, outweigh the objection." efficient to this historical record, as [Carroll (e.g., all window command, we mean menu into the global as in the Star environment). Julie objected because the currently active bar a window command a Kevin proposed the idea of incorporating them window, By rationale for this feature in the sense of historical record "The issue of where and pointed out namely the placement of at the top of the screen."* e.g. us take as an example a illustrate, let particular feature of the Macintosh operating system, commands 6 sense would be window commands, how it easier to namely how a given what are the tradeoffs it the other possible among them? some In our description of the they are related, and what often difficult to provide such a description in a systematic way, but an Since the unit of an artifact is often ambiguous (e.g., Macintosh, Macintosh OS, Macintosh window system, Macintosh menubar, or the position of the menubar), we will use the term artifact in the general sense to mean any feature, usually a small one, for which there is design rationale. Comparative Analysis of Design Rationale Representations example found is in [Card et ai. 1990; McKinlay et al. 1990], which provides 7 a vocabulary of the primitives and a set of composition operators for describing the design space of possible input This meaning of design rationale seems different from the devices. emphasis on design rationale not being emphasis not on a particular artifact a record first meaning but a construction, and from the second but on the relation among in its in its possible alternatives. Models of Design Rationale We now develop a series of progressively richer models of design rationale, which provide framework in which we can place the three different two meanings are discussed immediately below. The possibility space, is discussed shortly Design rationale in the way our it is. reasons as So in shown model we as a historical record and that led to the choice of as a set of claims an a first rationale, as a an designed the artifact is associated with a body of to extract and sequence On it The that led to the design, we justify the design of the artifact. menu is better because it The record at least of the process that artifact If was chosen. we wanted such a record and provides a basis a logical tells us the from which can represent the logical reasons directly, no matter how or in to i.e. what order they were artifact is an example because these claims These claims are logical also in the sense that the embodied by an context in which they are true have to be why rationale, respectively, an example. from the record; but the other hand, set of claims is The reasons can be artifact. artifact.^ one kind of reason the reasons justifying the choice of an artifact "The global is meanings of design embodied by an artifact tells us we would have infer the logical reasons. would why can give for an to the example about the window commands above historical circumstances articulated. The in Fig. 2a. roughly corresponding justification, an explanation of is historical or logical, free text meaning of design third of design rationale, an artifact There are different kinds of reasons that The rationale. after. most general sense first meanings of design made explicit. For example, a claim should not say leads to smaller implementation" if it really means "The We believe that the distinction between historical and logical reasons breaks if we push it too far because a purely historical record per se does not really give us a reason. It would give us a reason only to the extent that we can extract some logical structure out of it. Nevertheless, we believe that the distinction is useful for the purpose of evaluating representations because for a given representation we would like to know what it makes explicit and what we have to infer from it. Comparative Analysis of Design Rationale Representations (a) MODEL (b) MODEL 2: I : An artifact is assoaated with a body of Alternatives and their relations are all made the arguments relevant to the design of the artifact explicit and the arguments about individual ahematives can be differentiated. Ct^'^ ''^'k^lr^ '2.^*% a CTi) "^^3 Figure 2 (a)-(c). Progressively (continued on the next page) more differentiated models of design rationale Comparative Analysis of Design Rationale Representations ^'ne,^ (d) MODEL 4: Criteria used for evaluations further differentiated in the argument and their relations are made explicit and the arguments about them can be space. MODEL 5; Individual issues are made explicit, each of which contains the alternatives, evaluations, and criteria used in (e) discussing the issue. A part of the argument space includes the meta-arguments about the issues and their relations. Figure 2 (d)-(e). Progressively more differentiated models of design rationale (continued from previous page). global menu is implementation. better This in is memory." Extracting these the context of the Macintosh because important in a logical reasons system is like the it leads to smaller Macintosh which has a small not an easy task; once identified, however, they provide the advantages of being testable and general. Comparative Analysis of Design Rationale Representations The internal structure of these reasons can be they can be completely undifferentiated. we gave earlier as more relations such is an example of a we explicit, the natural language description that If we were make to the historical & and Replaces. An example Rosson 1991] that for describing the claims not historical is "embodied" in an In this representation, the claims themselves are represented in natural language, but the claims are grouped by the questions they answer how do do I We this? made more Qualifies, called a is historical record. Initiates, the representation used by [Carroll artifact. be At one extreme, explicit to different degrees. An example can differentiate further by making explicit such roles and relations Second-Motioner, as: Initiator, made 10 explicit what can I do, how does that work, and by providing such constructs as Logically-Implies, Supports, Denies, and Presupposes. body to: can also imagine a representation where the logical support relations can We Argument Space, will use the term. to refer to what we have of reasons, because the reasons are captured either as a historical record of the various arguments relevant for the design of the artifact or logical arguments underlying the design. There is much we can do with our model can help us answer first do today?" Such a representation can deal with this problem? & model the question, Can we design rationale, mentioned above, suggested in Smalltalk). which can be practice" of Our first tested and compiled to form, in the model, however, does not help very representation based on the model has itself we How how long run, to did other people [Carroll their representation of issues for redesigning much with Saying that cannot answer these questions. Of course, text, in detail issues can be this an artifact (the View couched as a design hypothesis, "a contextualized science out of interaction. qualify this statement immediately. language free many They discuss how these human computer based on week and what do we need from the past design cases? Carroll and Rosson Rosson 1991] provide a good example. They report Matcher last answer the questions: also help us learn A representation of design rationale. "What did we discuss if it the other questions, though does not help the user much is we should not to say that works hard enough, and as long enough information captured, even can answer these questions. So the real issue is in the we as the form of natural how much the model help us answer these questions either by helping us see the structure better or by enabling us to define computational services that help us answer the questions. differentiated the cost in models allow us to answer these questions more some other ways [Conklin & Yakemovic 1991]. easily, We will see how more although they increase Comparative Analysis of Design Rationale Representations 11 by making multiple alternatives and Our second model (Fig. 2b) differs relations explicit. Design involves formulating several alternatives, comparing them, and from the merging them, as many of the questions alternative made is our the ones that have been rejected. their attributes (e.g. current status among Once the alternatives many of the issues when this set of alternatives explicit, including become The comparing (e.g. explicit, we can talk '), about make one alternatives, other than the and the knowledge used finally chosen, are interesting in evaluating We situational constraints change. alternatives, displaying the even argue about whether an alternative them is because are useful in other contexts, for use the term. Alternative Space, to refer to multiple alternatives and their relations. These relations among the alternatives can also be may model, only a single first the alternatives explicit (e.g. specialization, historical precedence), alternatives that specialize this alternative), or worth considering. our such as "rejected" or 'waiting for more information define computational operations on them example, indicate. In list their and the multiple alternatives are present only explicit at a given time, argument space. Our second model make these implicitly in the the relations in first be not only the linear sequence that relations such as layers we Historical relations historical or logical. usually describe as versions, but also and contexts [Bobrow & The Goldstein 1980]. more complex logical relations may include Specializes, Generalizes, Elaborates, or Simplifies. Or alternatives can be related through a design space [McKinlay et al. 1990]. To the extent that represent these different alternatives and their relations, within the scope of the representation. space within its scope because one of its glBlS, for example, goals is 1988]. Position, with among The constructs we want a representation to say that the alternative space seems among these alternatives" [Conklin in glBlS for representing the alternative which we can describe multiple alternatives, space consist and the specialization & of: relation the Positions. By now, we have an alternative space connected to the argument space, as shovwi each of the alternatives, there are arguments describing the reasons for just as in our alternative, first i.e., its model there are arguments describing the evaluation that it was chosen. Some argument can support an of the alternative while it is For status of that single arguments can be shared; denying another; so in Fig. 2b. current evaluation, for example, an better to think of the arguments about the different alternatives forming a single large argument space, as shown Fig. 2b. is to include the alternative "to capture alternative resolutions (including those which are later rejected), [and] trade-off analysis Begeman we in Comparative Analysis of Design Rationale Representations Once the alternative space us answer some what are their pros represented, is of the questions in our we can imagine how we can make a system that help To answer "What are list. the alternative designs and cons?," we can associate an argument space with each week?" consider this alternative, and we need some to how the representation of the alternative space, related to the we week?" consider this alternative, and to among Part-Of), can make a system help us are the alternative designs how is use some historical relations related to the it (e.g. Replaces) To answer "Why do explicit multiple alternatives, what is argument space namely the chosen one, Similarly, the about the arguments why generalize, one that we discussed last or structural relations (e.g. Is-A- the alternatives. However, once we make the last With and cons?," we can associate an argument space with each of the their pros we need discussed how we alternatives through the links such as Supports or Objects-To, as in gIBIS. we even we the alternatives. To answer "What can imagine in Section 2. one that among such as Rqjlaces, historical relations, answer some of the questions posed and what are is it and of the To answer "Why do alternatives through the links such as Supports or Objects-To, as in gIBIS. we even 12 (Fig. 2b). we need In our first argument space contained reasons for the other alternatives are about carefully single artifact, for the choice of that artifact. why they have their particular evaluation status, "Waiting for More Information", "Rejected". more to articulate model when we had a they were not chosen, or, to e.g. "Still in Consideration", These evaluation status could be nominal categories (such as the above examples), ordinal categories (such as "Very Good," "Good," and "Poor") or a continuous measure (such as the probability that the alternative will achieve a given set of goals). Therefore, explicit we and introduce the evaluation space (Fig. related. evaluation measures Usually, we we do "Very Good," "Good," "Poor," "Very Poor") them to manages these produce a higher values mean. values, for level if is sufficient However, example if the evaluation status are when we we want among leave it for the to define to made among the these values (e.g. human user to any computational that automatically propagates summary, then we need and merges be very careful about what these We need to specify the units of measurement, a calculus for combining them, and a model specifying what they mean. Even example, where Often, the implicit ordinal relation use. assign these values to the alternatives. service that 2c), not and need not specify any elaborate relation the human summary measure, user then interpretation does not is we need become in the case where these actions are left to human, for expected to combine these values to produce a higher level to arbitrary. set down what these values mean so that their Comparative Analysis of Design Rationale Representations Making the evaluation space explicit allows us to differentiate space: (1) arguments about why an alternative has arguments about the alternatives themselves, an object as an alternative or whether That alternative. is, differentiated in the now answer evaluations?" We "What are as: Our models so far can also explain is duplicate do not make For example, menu the how an two most favorable evaluation items" is it explicit the criteria and used in we know can group If that the "Reducing Screen Clutter" we have is a to this criterion or criteria way Menu to at the top of the used as a criterion for evaluation. becomes Knowing how less important, then this goal (for example, this criterion is related to others (e.g. of achieving "Easy to Use"), also allows us to assign proper change its importance, when the related criteria changes. We space to refer to these criteria and their relations. As Fig. 2d shows, once the criteria space explicit, it is is to represent the arguments that appeal to this criterion arguments that presuppose we can grouping those arguments which are about the Hence, all important argument "We do not need the criterion changes or to all the these arguments less important). use the term arguments this particular producing an evaluation. However, a pro-argument for the alternative "Global this criterion explicit, importance to the and explaining how their relations are usually quite important to is can do appropriate things making and "Sun the release change our was made by pointing artifact in question, and weigh them against one another. we alternatives so far?" How would server. Screen" because o/the goal of reducing screen clutter, which By making another derived or computed from them or related to other measures. the criteria used for the evaluation explicitly. (2) should or should not even consider these different kinds of arguments can be 2c, X/NeWs argument space behind the evaluation measure why we this alternative is really a special case of in Fig. released their just two components of the argument current evaluation status, and its argument space. With the representation of the evaluation space, we can questions such Microsystems in the shown as e.g., 13 and their relations. important that a language whose scope includes the different attributes of the criteria argument space by further differentiate the criteria criteria and the relationship among them. space represent the For example, it should allow us to represent the importance of these criteria and the synergistic or tradeoff relations among them. A set of criteria can be sub-criteria of another in the sense that satisfying them facilitates the satisfaction of the latter. These sub-criteria can be related among themselves in various ways. They can be mutually exclusive in the sense that satisfying one makes it impossible to satisfy others. They can be independent of each other in the sense that satisfying one does not change the likelihood of satisfying others. These sub-criteria can be related to Comparative Analysis of Design Rationale Representations ways their parent criterion in various satisfying With the of all them answer the questions such important anyway? as: we what if can now we do The answer might be evaluation of the alternative X were based on in the sense that equivalent to satisfying the parent. is space represented, criteria They can be exhaustive as well. 14 X changes how see the system might be able to help us not consider portability? "If we "High" because to why or is portability give up the goal of portability, then the these claims that argue against all the importance of portability." or "Portability is important because it is a subgoal of another important goal, Have a wide distribution." These answers can be derived from a representation criteria, the representation if these questions. It is not obvious how explicit the relation space seems a necessary condition if criteria is between evaluations, not sufficient for answering these questions can be answered even space are represented explicitly. criteria makes and arguments. Of course, representation of the we However, the some if parts of the explicit representation of the criteria are to answer these questions. At least, we would have the information necessary to define an operation that will give or suggest the answers to these We will give some examples questions. So we have far, artifact, and discussed the structure of a identified namely which of such operations later in Section 5. of the alternative designs should representation of such local structures alone, namely criteria space, we unresolved issues and what are depend on its we questions, A decision often individual issues are related. requires we need A decision can be a specialization of another case of the second. term Issue Space has as its them. capture to A how we can with this still the latter requires the some questions problem? Can first is, decision as we that learn a more the first detailed and we use the Once we have an issue as an and "Actions Taken" with issues and what are we currently among the issues will allow us to "What are the unresolved we have is decisions. making therefore, a single decision that in Fig. 2e. Representing the dependency relation answer the question "What are the issues that depend There are if and are the are the issues that these decisions are related, shown What more global picture of how associate the attributes such as "Status" and answer questions such doing about them?" if unit in this issue space internal structure the other spaces, as explicit element, issues important It is to refer to as: and /or influences many other For example, a decision can be a sub-decision of another decision. a alternative space, such What currently doing about them? To answer such this issue? in the list underlying an However, with the choose? argument space, cannot ask some of the questions still single decision we this issue?" as: How did other people deal We argue, however, that the five not yet covered such from the past design cases? Comparative Analysis of Design Rationale Representations spaces so far identified, the spaces of arguments, alternatives, evaluations, can contain enough information to answer these questions. In Section 5, 15 criteria, and issue, we discuss computational operations that help us answer these additional questions by exploiting the structure of the five spaces. EXISTING REPRESENTATIONS 4. There are only a handful of representations that have been used Kunz & rationale [Toulmin 1958; 1988; Potts & Rittel 1970; for representing Marshall 1987; McCall 1987; Conklin Bruns 1988; Lee 1990a; MacLean et al. 1991]. Most, though not design & Begeman of these all, representations are heavily influenced by the IBIS structure for representing issues [Kunz Rittel 1970] or MacLean propose a representation based on 91] also In the next three subsections (4.1-4.3), rationales. detail by Toulmin's model of argumentation [Toulmin by defining we In Section 5, with constructing design discuss these three representations in we present a language, called of the limitations of the existing languages that Some qualifications are in order before we them representations, but only to evaluate is proceed. DRL, which we believe overcomes we discuss Our in this section. intention is not to crihcize the existing design rationale representation language. Even as a mainly with respect to their expressiveness. We would like to emphasize again that the adequacy of a representation can be evaluated only with respect set of tasks. be able it to One representation do more with it. In discuss a few other studies that bear on design rationale many then, our evaluation & et al. 89; scope and evaluating them within the framework outlined above. their the next subsection (4.4), representation. their experiences we [MacLean 1958]. may be However, if much more what it expressive than another; as a result enables is to a we may not in the set of desirable tasks, or if enables those only by sacrificing other more important constraints, then the additional expressiveness is not worth what representations allow or it gives us. do not allow us What we to do; discuss below We do not intend to is only what the existing make value judgments about whether the representations should or should not do. Some of these languages were designed with different goals, which in turn determine the tradeoffs adopted in the designs. Therefore, the following discussion should not be construed as a criticism of these representations, but only as an articulation of their scope and languages with respect to expressiveness. We clear that also want information, their adequacy as design rationale representation whenever we say some to make it does not mean that people cannot infer that information from the that a representation cannot express " Comparative Analysis of Design Rationale Representations Take representation. example in the language representation of design rationale that a natural beginning of the paper. If we been ever expressed if we work cannot express some information, that make that makes amenable an can always retrieve the information When we say that a representation way that help people see the structure better or computational manipulation. to IBIS (Issue Based Information System) 4.1. The IBIS structure was originally developed in representing designers' argumentation activities. [Kunz & used by gIBIS [Conklin Since gIBIS & Begeman incarnation of the original IBIS discuss one used by them The goal among , "a & we some One variations of variation is purpose of it have been the representation hypertext tool for exploratory policy discussion. Yakemovic and can be regarded as 1991] a modem use gIBIS as the context for discussing the IBIS Other variations include PHI (Procedural Hierarchy of Issues) [McCall 1987] representation.^ the 1988], most well-known [Conklin is Rittel 1970] for the Since then, used by a few systems for representing design rationale. and as that the representation does not provide constructs the information explicit in such a it we hard enough. we mean we gave keep a detailed enough record of what happened, or even a video recording of the whole design process, that has 16 [Potts & Bruns 1988] for the rationale module in their representation. We briefly in Section 4.4. of gIBIS is to capture "the design problems, alternative resolutions, tradeoff analysis these alternatives, and the tentative and firm commitments that were made in the process of the decision making process". Figure 3 shows the objects and relations that form the language of gIBIS and Figure 4 shows an example representation. In gIBIS, one raises an Issue such as where to "in the global menubar put the window commands. Arguments can be created need to duplicate "Requires long in the figure. to and "Commands Support or Object-to a Position. commands mouse Positions are created to Responds-to the issue (e.g., at the top of the screen.", for each travel." objects to window it. " supports the first position, and the argument Also, an Issue can be related to other objects as The gIBIS model extends the generalize/specialize relation at the top of each window.") For example, the argument "Don't among Arguments original IBIS as well as shown model by introducing: the among Positions, an additional We believe that the distinction between historical and logical reasons breaks if we push it too far because a purely historical record per se does not really give us a reason. It would give us a reason only to the extent that we can extract some logical structure out of it. Nevertheless, we believe that the distinction is useful for the purpose of evaluating representations because for a given representation we would like to know what it makes explicit and what we have to infer from it. Comparative Analysis of Design Rationale Representations 17 C_J_^> Figure f Commands I top of the at th^l window. Commands Ilside ot the at 3. The glBIS vocabulary. thd window C ^" - ^ ^ /"•" responds-to responds-to _J Dimmed commands by Figure 4 . their still distracting mere presence. An example V 'I (^N°J— I in IBIS. commanas rea jistjacting? Comparative Analysis of Design Rationale Representations 18 Other type which serve as an "escape" mechanism for representing things not representable with the available constructs, and an External type for external objects except when we to both IBIS and gIBIS discuss the generalize/specialize relation for Position and Argument. The scope of the gIBIS representation depends on what an issue sense general such as documents or common sketches. In the following, our discussion refers to the structure mean any question to takes that If is. a we of set take an issue in a very Positions, then the Issue/Position/Argument structure can represent a fairly large part of the design rationale spaces. The internal structure of these spaces, however, is not well differentiated in gIBIS, as we discuss below. The alternative space is represented in gIBIS by Positions and the among relation them. Since multiple Positions can be created for a given issue, gIBIS allows the representation of multiple alternatives, thus offering at least the richness of our among the Positions that we can represent in gIBIS although there are other relations that can connect Questions or ls-Suggested-By)7 model is 2. The only relation, however, the Specializes/Generalizes relation, a Position to objects of other types (e.g. Thus, gIBIS can answer the question "What are the other alternatives being considered?" but not the question "How are these alternatives related?" unless they happened to be related via the Specializes relation. The unit of the argument space in gIBIS applicable to the current Position (e.g. "In the global position is example, or we is Argument. An Argument only among the if (e.g. "The commands no construct, however, for qualifying Argument "the global menu containing menu all Argument reduces the screen possible action items itself" is valid. the arguments is again limited to the Specializes relation. "Irrelevant menus can be dimmed" Argument, "The commands not applicable distracting." To express this relationship about whether the original Argument argument." However, the an argument. For This limitation implies that an Argumentcannot Support or Object-to another Argument directly. the why at the top of the screen"), indicating a reason is not be shown and distracting.") can support or object to a cannot indicate that the existing Argument, "the global does not become so big relation menubar not a good one. There clutter", is valid The is window would I was not is is an argument objecting to the current among Arguments window would in gIBIS, we have right or not, create Positions "yes" able to find any discussion of it yet. For example, to the original be shown and to create and "no", an Issue and then Comparative Analysis of Design Rationale Representations A argue about these Positions.^ new to create three Issue, a A more interface that hides the intermediate details. that new objects: a However, we might be able supporting the "No" Position. is We this claim." it might try to leads to the Argument serious limitation of this representation me the claims that respond to all and the issues link responding claims. However, that contain the too general for that purpose because is tin reduce the complexity with an answer these questions by following the Is-Suggested-By that connects the original claim the Is-Suggested-By link "No" Position, and to does not help us answer questions such as "Show it that is For example, in order to just point out the factual inaccuracy of a claim, proliferation of Issues. we would have possible criticism of this representation 19 does not allow us to it distinguish the issues that contain responding arguments from those that do not. In gIBIS we we cannot argue about relational claims. For example, there agree with A and B, you can argue about. but not that A To with another example, suppose illustrate whether portability should be a only way we can do a criteria?", we so in IBIS criterion for evaluating the by creating and suggest possible answers should not". Then, the original issue make is we by an Is-Suggested-By explicit the relation between the object Being able want to to new in the form of relation. we can argue and the agree with However, We criteria A commands; he may point out irrelevant commands. to when is is a that the global The user may want is to new issue to If we make important for other reasons as well. People often to make that A supports B. For bad idea and that seeing irrelevant it first. That distracts the user menubar does not have is, one may by showing to show the the distinction clear because they do not even more crucial be construed as denying either of the claims alone. The distinction the representation The again does not being questioned. and B but not with the claim bad idea but not because irrelevant want argue about should" or "No, this this representation distracting but not that the second claim supports the a we can also relate commands is to given issue. directly about the relation, Is-a-Criteria-for, agree that the global menubar menubar not an object that consider portability as positions, say "Yes, may is for a we example, one believe that the global is original issue. argue about relational claims may no way of saying that we want posiHons say "Should Issue, between the new issue and the portability say that they a can argue about these positions. a relation an explicit object, then is Supports B because a link in gIBIS is be used to provide computational services. Mixing denial of node is to show the first argument as supporting the position that the second objects instead of objecting to the second argument directly because an argument cannot respond to another argument in gIBIS. This representation does not make explicit the relation between the two Another alternative to, arguments. Probably for that reason, gIBIS does not encourage personal communication) this representation (Conklin, Comparative Analysis of Design Rationale Representations with denial of link this just is not right semantically. you define a computational service based on type of careless semantics, such as evaluation management that propagates and merges evaluations, what you get would not make much The evaluation space used by gIBIS "Chosen" assigned Information," but beyond that try to derive some space criteria gIBIS is we We could it is We some nominal among difficult to more sophisticated evaluation measures can the positions based imagine what the calculus would be gIBIS. gIBIS provides we cannot argue about them; nor can criteria. Again, we we Fig. 4 indicate we no vocabulary for describing the shows an explicit the relation any relationship, such as mutually exclusiveness, among the between this issue But, again, this representation does not and the criterion in question. be even more difficult for a For example, this relationship; no easy way disagreements to For Issues. Issue ("Should small implementation be a goal?") related to our representation of the criteria entails other limitations. people to see Since criteria are cannot represent the reasons for having these could indirectly represent these relations by creating additional original Issue via an Is-Suggested-by link. As we arguments, and like. a serious limitation for a design rationale representation language. criteria; is to the on supporting or objecting use in evaluating the alternatives. This lack of explicit representation of criteria in not explicit, example. and categories such as "Rejected" use finer categories, such as "Waiting for More could assign uncertainty measures beyond the scope of is sense. not clear what it is ordinal ranking arguments. However, criteria consists of to the Positions. be defined on the gIBIS structure. The If 20 it will accommodate the changes. among people because It is more It will make difficult for machine. Lack of explicit when goals change, there difficult to the criteria they use in their be isolate the real arguments remain will see in the next section, explicit representation of goals implicit. can provide modular representation of arguments, multiple viewpoints, and a basis for relevance matching. In glBlS, the unit of the issue space is describing the relations about issues. and Suggest another window commands Issue. Issue, Issue can Generalize, Specialize, Replace, Question, Issue can also Question a Position or among we command pop-up menu should should have small implementation as a goal. be An an Argument; an Issue can be Suggested-by a Position or by These relations are quite important the issues. for Figure 4 shows that the original Issue about the placement of suggests other Issues: whether the considered as an alternative and whether an Argument. and gIBIS provides several constructs an An in describing a more global relationship In particular, as an issue or a decision often gets reformulated and Comparative Analysis of Design Rationale Representations differentiated, the relations such as Replace however, if we can somehow show whether and seem Specialize a given set of relations essential. is 21 It would be nice, complete or adequate. Toulmin's Model of Argumentation 4.2. Stephen Toulmin, a philosopher, proposed many has been adopted for [Bimbaum et al. 1980; model a of argument in 1958. Since then, the model purposes, including the computational representation of arguments Lowe 1986; Marshall 1987; Streitz et al. Figure 6 shows an 1989]. example Toulmin representation of an argument.^ CLAIM Placing window commands in the global menu reduces commands need not be usually duplicated for X each window screen clutter. unless No duplication saves space. tlie global containing menu all possible commands becomes 5 on-account-of Common so big Sense Knowledge Figure A Claim menu is the 5. An example of an argument in Toulmin's window commands in the global support the Claim by producing a Datum, "Commands main assertion being made reduces the screen need not be duplicated clutter. for by producing a Warrant, ") We (e.g., each window."^^ e.g. "If representation. "Placing the We justify the leap you need not duplicate from the Datum things, that saves space." usually an inference rule or a principle that licenses the inference from Warrant is not often absolute; so Datum inference from the We will use the interested in the if to the we want Claim, to we express how Datum to the A Claim Warrant is A to Claim. strongly this Warrant supports the use a Qualifier. In our example, the Qualifier term Toulmin's representation rather than Toulmin's model because to the extent it has been or can be used as a representation. we is are model As pointed out in [Newman & Marshall 90], Toulmin wants the datum to be a singular fact, such as something that you can point to, because we think he still wants to retain the syllogistic form, where one gets the conclusion by having as the minor premise, a singular fact. So he seems shll under the bound of syllogism, at least at the time of his writing, despite his rebellion against it. The datum of our example is a more general statement. But we use this example first because the distinction itself between a singular fact and a universal statement has been called into question in philosophy and because in design it would be unusual to support a claim always with a singular fact. Comparative Analysis of Design Rationale Representations which "usually", have sum be larger than the to question, then we appeal In through Rebuttal fails i.e., a our example, body we of of the called the argument space. The goal There supported. how Toulmin structure is no attempt they are related. As such, module elements of design rationale. Toulmin was that of facts, tell why us called into is we in general or legal precedents. Toulmin structure is to make limited to is how explicit what we a claim is any evaluation measure, or to represent alternatives, goals, we be shared might to a Warrant as a design rationale representation evaluate Toulmin's representation only as a representation argument space, which might for the knowledge If ') appeal to our commonsense knowledge. Other types of backing might be physical laws, historical The scope indicate the excepting "The object (e.g., of these objects locally existing. to a Backing, should accept the warrant. We the strength of the inference. qualifies circumstances where the Warrant 22 Even to delineate the logical then, we interface with representations for the other should keep in mind that the original goal of support structure of an argument, not necessarily provide to a representation to be used dynamically for capturing design rationale. As argument space of design a representation for the we For example, menu that the global possible only command way to do so is (let us assume that this is problem by It is much all the of denying a Claim; Even with such explicitly to menus were menu items" as the as supporting and in construct, such naming any "A does not reduce representation [Bimbaum et al. people corresponding construct. Claim itself is We B", is it would would slow down 1980] Datum and Claim have solved and as Supports So Toulmin's representation should be extended at claims. we want to qualify a We can qualify a Warrant Claim or a Datum, there no is could of course build into the Claim any qualification so that the of the form, "Global not too big," but then if and Claim. without some construct that also not easy to qualify a claim using Toulmin's representation. using Qualifier and Rebuttal, but for this other, the denial relationship awkward the original link between accommodate denying case, Datum all The invented). new introducing another link called Attacks. least this deny to has to contain clutter" has to contain human comprehension or machine computation. Some this limitations. menu two Claims are the negation of each all. menu many you want by creating a Claim "the global menu does not reduce the screen awkward way indicates that the negation. Suppose before pull-down Whether denying "A reduces B" means the same not be explicit at its reduces the screen clutter because the global items presents "the global certainly an rationale, Toulmin's has can only deny a Claim by supporting we might menu reduces screen clutter provided that the global menu as well use a natural language as our representation. is The Comparative Analysis of Design Rationale Representations purpose of making these components distinct to is 23 allow us to see the relations more clearly, to formally manipulate these components and relations, and to incrementally construct pieces of the argument. Toulmin's representation also suffers from building into object types context-dependent such as warrant or backing. The difference between warrant and backing the object itself but only in the fact that backing is its we context-dependent role [Marshall 1987; Streitz et in some other way, we show as Most if we want to al. 1989; Newman & support a Datum change So context. a separate type called Backing. its role to Claim, instantiate the now without Backing) schema, and use a Datum about Warrant that we need we first change fill it in. we want If to However, then (that back the Claim? a separate type for its latter is, It is the it is to Claim, not clear why why not Toulmin structure not clear what when we do support For approach. from Datum role its support or deny a Warrant, four-components to an object type should in Section 5, or of the Toulmin-based representations Marshall 1990] take the object, then instantiate the five-component schema, and we need Datum, depending on the assign a type to an object based on the inherent properties of the object and indicate dynamically change depending on the context. example, roles, not in the nature of something that supports warrant. The same object can be a Warrant or a Backing or even Claim or either is is special not need a separate type to support other things such as a Datum. We understand that Toulmin's reason for introducing Backing as a separate type is to category for the kind of knowledge that common sense knowledge, or physical laws. must be something that it is is provide a rarely called into question, such as the written law, We agree that for an argument to be resolved, there that the involved parties can agree and appeal However, we believe to. misleading to represent such knowledge as a distinct component of a rigid structure such as Toulmin's. For one thing, the five-component schema leads people argument has to make explicit all these components to be complete. people to unnecessarily represent the obvious; for example many to believe that warrants, when explicitly represented, are often as obvious as the one in our example, "No duplication saves space". often in the process, people are led to argue about whether something rather than about the substance of the argument. leads people to believe that having five components is if Warrants, or chain of Warrants before it is five we Also a warrant or backing, components, because there might have better to define the representation in is Conversely, the closed nature of the schema an argument has these arbitrary, an This belief often forces get to something that an open-ended way to we it is complete. But, be multiple Data, multiple agree on. that allows as We many believe that objects to be created as needed without necessarily specifying the boundary of an argument and without Comparative Analysis of Design Rationale Representations requiring the obvious to be represented unless it needs to be, e.g., unless it is 24 called into question or argued about. There are numerous other limitations with Toulmin's representation, as [Newman & Marshall 1990] point out from their experiences of using Toulmin's for representing legal arguments. For example, they had to make Rebuttal have at least four additional meanings what different types of objections, each corresponding to is being objected these different types of objections should explicitly indicate can do so in a graceful way, as we hope to show what they to. to represent the We object to, believe that and that we in Section 5. QOC (Question, Option, and Criteria) 4.3. QOC is rationale. a representation proposed by Design rationale in QOC is [MacLean et al. 1989, 1991] for "constructing" a co-product of design that has to be constructed alongside the artifact makes QOC different from, say gIBIS, whose goal DATA: commands not applicable to current window would be shown and distracting DATA I In global al menubar lop of screen On each window Positive evaluation Negative evaluation itself. is This emphasis at least includes capturing the rationale as unfolds. I design said to be not a record of the design process, but ir\stead it Comparative Analysis The major constructs this paper. QOC is of QOC are straightforward The a Question. and map unit of the alternative space to Issues and Positions the criteria space with Criteria. "bridging criteria" (e.g. Design Rationale Representations clearly to the 25 framework proposed in Figure 6 shows an example represented in QOC.^^ The unit of the issue space in roughly correspond one of Easy a if it is The to use). corresponding to more A Criteria (e.g. specific one Questions and Options an Option. is However, unlike gIBIS, in gIBIS. QOC Reduce Screen Clutter) that derives its can represent said to be a is from a more general justification units of evaluation space are links labelled with "+" whether an option does or does not achieve a given representing the argument space are Data, Theory, "-," and The constructs criterion. One supports and Mini-Theory. for the evaluation ("+" or "-") of an option with respect to a criterion by appealing to empirical Data "The mouse (e.g. there is is a Fitt's no relevant data Law at Device") or to an accepted Theory hand or construct a Mini-Theory, which [MacLean existing theory to is draw (e.g. "Fitt's Law"). on, the designers When may have to an approximate explanation of part of the domain. provide an illuminating discussion of the other forms of justifications for et al. 1991] design, such as various forms of dependencies and metaphors, no specific constructs are discussed for representing them. QOC as we understand argument it has a number of limitations as a representation language. First, in the space, constructs like Data, Theory, or aspects of arguments. For example, connmands can be dimmed" should be nor a theory. Nor in theories. is it clear it is Mini-Theory not clear treated given that whether and how In the alternative space, there is we can seems to treated as a fixed type, rather ir\flexibility. but is That it is it. dependency, but no In the criteria space. Bridging Criterion role into a being a Bridging Criterion is fixed type results in unnecessary not a property inherent in the object a relation that the object has to another criterion. has with another criterion. many as "Irrelevant neither a piece of empirical data Otherwise, we have itself As such, one should not have classify a given object as a Bridging Criterion but instead indicate it capture than a role that a criterion plays with respect to However, building a is, to argue about theories, or individual claims a reference to cross-option specific constructs are discussed for representing another criterion. do not seem how an argument such its role to unnecessarily to through the relation change object types depending on which object we focus on. The to 1^ QOC constructs for the evaluation space are "+" and "-" links. These evaluations are said be supported by appealing to empirical data and accepted theories. We could represent only a part of our example because rest in QOC. it was not It is clear to us not clear, however. how to represent the Comparative Analysis how we of Design Rationale Representations can differentiate those arguments that support the overall evaluation from those that deny or qualify Suppose it. "Commands can be found that the empirical Data, positions, "produces the positive evaluation of the Option, "The global menubar screen," with respect to the Criterion "Can easily access commands."' there is another argument same respect to the create 26 criterion. two evaluation The is. links, is mouse Suppose, however, that option with travel.") against the can represent the design rationale in one of two ways: one labelled + and the other -, (1) between the same Option/Criterion how option has the disadvantage of not differentiating explicitly latter each datum contributes evaluation link We requires longer only one link labelled with whatever the net effect of these pro and con pair, or (2) create arguments (e.g. "It in fixed at top of the to the net evaluation since the link between any Data The not labeled (such as with "Supports" or "Objects to"). keeping two evaluation links avoids problem, but this we first and QOC are not sure whether the option of allows multiple evaluation links from one Criterion to one Option. Other Representations 4.4. In this section, otherwise, we discuss a few other representations, explicitly for design rationale or whose scope includes the aspects PHI (Procedural Hierarchy we discussed above. of Issues) [McCall 1987] overcomes gIBIS by allowing a quasi-hierarchical structure semantics of the hierarchical relation an Issue, A, is resolving B. In the the more first is a (i.e. issues, if A "serves" B second ~ for it a child sharing nodes (i.e. node Hence, unlike of the in gIBIS, is is, Answer In the issue space, if is A resolving helps a child of another a child of another can respond Argument. Furthermore, PHI argument space On the other hand, the expressiveness of is extended over that is to if the if a is first is a an Argument directly by i/uasi-hierarchical and allows in gIBIS. PHI is limited in the following aspects. to express the specialization of issues in specialization, we that multiple parents) and cyclic structures. This way, the expressive power for the how of the limitations with example, a mechanical system subanswer of a system. In the argument space, an Argument making -- the alternative space), an specific version of the response to the second. many answers and arguments. The different for the different spaces. is a child of another Issue, B, answer space among namely the By the same token, it is PHI because hierarchical relation, has been not clear how preempted by the serve to represent the serve relation in the because the hierarchical relation in the answer space is It is not clear the natural representation of relation. answer space the specialization relation. That is. Comparative Analysis of Design Rationale Representations how do you (e.g. Use represent an Answer that serves another Anszver in the sense that adopting the a large screen) will facilitate adopting the second (e.g. also inherits many make claims, for example, that A [Fischer et components, e.g. 1989] al. CRACK, the criteria explicit, Supports is B. an attempt interesting as argument to make the criteria space explicit. domain about kitchen design. The other component, ViewPoints, uses PHI make. JANUS integrates the cannot relational One bridge two representations. uses a rule-based language for representing for the decisions they we space: and cannot argue about the Again, PHI does not first Reduce Screen Clutter)? PHI of the limitations of IBIS with regard to the qualify a claim, cannot JANUS 27 specific of its knowledge, to represent the rationale two representations by finding the appropriate rationales represented in PHI for the particular issue that designers face in the construction phase, that while using is CRACK. Although the current interface is limited to that of locating the relevant parts of the representations, bridging a design rationale representation and a domain representation is a very important topic of research because such a bridge can allow us to represent the relations domain specific & [Potts Fig. 7. among the alternatives or the criteria in Bruns 1988] outlines a generic model The model extends model the IBIS One design including a software design. formatter, associates with it for representing design deliberations, shown to represent the derivation history of starts and so on until we make each Alternative may be a is the plan concrete justification essentially an IBIS structure, representation as well. Artifacts, and making better. logical relationship rationale, explicit, it The enough the plan more On among its path, this to to a of what we more concrete be implemented. Associated with showm in Fig. 7, said about IBIS applies to this series of progressively more concrete model represents the alternative space and the other hand, this representation the alternatives. in artifact concrete, associates internal structure of rationale, as much However, by describing a Justifications for argument space and an with an abstract Artifact, such as a plan for a the Issues that arise in with each of the Issues the Alternatives considered, some of which lead plan, more ways. still its does not represent the Relying on the IBIS structure for representing its shares the problems of IBIS such as: the criteria used for justification are not an argument cannot respond to another, and we cannot argue about relational claims. Comparative Analysis of Design Rationale Representations J 28 Comparative Analysis of Design Rationale Representations however, represent argument structures typically found in academic papers and to serve as to is an underlying structure arguments in for aiding "reasoned discourse ARL ARL would different structures in different contexts? may questions 5. we especially writing. The structure of the from the structure of the arguments not be a good design rationale representation language. raises interesting questions grained are the distinctions ", sufficiently different does not provide any construct ARL However, seems in design deliberation academic papers that Also, 29 for representing goals, alternatives, or evaluations. about the domain of arguments: Or are willing to Do arguments have how are the differences only a matter of make fine- Studying these in the different contexts? help us better understand the nature of arguments in general. DRL (DECISION REPRESENTATION LANGUAGE) 5.1. Introduction DRL [Lee 1990a] a is language that we have developed for representing and managing the qualitative elements of decision making: for example, the alternatives being considered, their current evaluations, the arguments responsible for producing these evaluations, and the criteria We used for the evaluations. elements, and environment we is the representation of these qualitative management system a system that provides an and computational services using a subset of design rationale in that aspects of design rationale. how rationale for capturing decision rationale rationale in our sense about call decision rationale call a decision it For example, a design rationale to generate the alternative designs. The scope it. Decision does not represent some important of may DRL, include the deliberations at least for now, does not include the representation of such deliberations. ^^ jhe exact relation between decision and design rationale that DRL and that is it rationale, however, has yet overcomes many of the limitations simple enough for the user. In this representation language, and point out ^2 to be articulated. Nevertheless, we believe the most expressive language that has been used for representing design rationale its in the existing languages in a DRL section, we evaluate strengths and limitations as such. way that is still as a design rationale does not help us generate alternative designs. It does in a couple of ways. argue about the existing alternative designs, thus helping them to see more clearly their strengths and the limitations. It also helps people retrieve the past decisions that contain useful alternative designs which are still useful for the current decision or those that can be That is DRL allows people not to say that so adapted. the new DRL to Furthermore, alternative that design alternative is still DRL can represent the relationship between the existing alternative and derived from it. However, being able to generate a new how it was generated may have been different from representing the rationales for Comparative Analysis of Design Rationale Representations 5.2. 30 DRL Description of Figure 8 shows the objects and the relations that form the vocabulary of DRL. The argument types for a relation are DRL Relations, for shown inside the parenthesis following the between them graphically. relations we example, are subclasses of Claim, as represented in DRL. We is not a description of the of the relation. All Figure 9 shows the Figure 10 shows an example decision rationale DRL describe the basic features of briefly, they allow us to represent the five spaces of design rationale. following name explain below. way that the user and discuss One should would use DRL. The in detail how note that the actual interface is briefly described later. A qualification is in order before we proceed further. DRL is an extensible language, and have a developed a method, called Dependency Analysis, that can be used add additional constructs below are what we given set of tasks. Therefore, the constructs of would be For example, if specialized to would be important to to more among presented dependency analysis method how dependency the requirements goals such as Replaces, Elaborates, and Merges represent explicitly. Unfortunately, we cannot discuss here, for want of that allows us to identify provide and specify their semantics precisely. snnn task-specific constructs using a task involves looking at the history of evolved over time, then the relations space, the DRL believe are essential for the generic task of decision making. For a given set of tasks, these constructs analysis. for a we to systematically We hope which additional constructs to describe it in another paper 13 Very briefly, however, dependency analysis works as follows. Given a task, we generate all of its to form a task hierarchy. From the leaves of the hierarchy, we identify the subtasks recursively objects and the attributes that generate a set of tables. change correspond relation. to a relation, we need. For all possible pairs of the attributes, we can systematically Each of these tables corresponds to the influence that a particular type of can exert on the other attribute. A subset of the cells in a table and the type of change signified by the cells in the set gives semantics to the in the first attribute Comparative Analysis of Design Rationale Representations Decision Proble Achieves (altemarive, goal) Supports (claim, claim) Denies (claim, claim) Presupposes (claim, claim) Is a Subgoal of (goal, goal) Is a Good Alternative for Subdecision of (dec. prob., dec. prob.) Suggests (object, object) Comments Answers (claim, object) (claim, question) k an Ansvi-ermg Procedure for (procedure, question) a Subprocedure of (procedure, procedure) Is a Result of (claim, Is Stronger Than (object, object, attribute) Is [ Is a (alternative, dec. prob.) procedure) Question \ Croup Wiewpoint (procedure \ Procedure Description Figure 8. Is Equal To (object, object, attribute) Tradeoff (object, object, attribute) k Executable Procedure The a DRL Kind of (object, object) vocabulary. ^o^ is a subgoal of I denies I Claim K^^^ is a good alternative for supports y |aaim[ Figure 9. The structure of a DRL decision graph. I laaiml Comparative Analysis of Design Rationale Representations An example Figure 10. A Decision Problem represents the the problem window commands. An Alternative at the top of the screen." alternatives: e.g. A "Minimize screen clutter". We menu is A clutter". elaborated to Every relation in Goal where to place menu DRL in Fig. is elaborated further in terms of is mean "Can easily access command a subclass of Claim, as shown its items" and in Fig. 8. For 10 represents the Claim that the Alternative, at the top of the screen," achieves the Goal "Reduce screen clutter." evaluate an Alternative with respect to a Goal by arguing about the Achieves relation between the Alternative and the Goal, We that requires a decision; for example, represents an option being considered: e.g. "global example, the rightmost Achieves link "the global decision graph. Coal represents a desirable state or property used for comparing the subgoals: e.g. "Easy to use" "Minimize screen 32 i.e., the claim that the Alternative achieves the Goal. argue about a Claim by producing other Claims that Support or Deny the Claim or by qualifying the Claim by pointing out the Claims that it Presupposes. Each Claim has the following attributes: "Evaluation," "Plausibility," and "Degree." The evaluation of a Claim, Comparative Analysis of Design Rationale Representations represented by the value of and "Degree" true, "Evaluation" attribute, its The attribute values. and the "Degree" of a Claim tells us to v^hat extent the goal true. and the Goal tells The overall us tells it is question. in a function of both of "Plausibility" of a Claim the Achieves link between the Alternative alternative achieves is 33 how its "Plausibility" probable the claim is For example, the "Degree" of what degree us to that the evaluation of an alternative is represented by the "Degree" attribute value of the Is-a-Good-Alternative-for link between the Alternative and the Decision Problem, resolution for the issue. This degree i.e., is Not the Alternative to the different Goals. evaluation. For example, we might the claim that the alternative is good alternative a a function of the degrees of the Achieves claims that link all of the three attributes have to be used for the require that a Claim be entered only if its "Plausibility" above a certain threshold, and ignore the "Plausibility" once the Claim has been entered. In case, we attributes is that can do away with the "Plausibility" attribute, and the "Evaluation" and the "Degree" become synonymous. There are other auxiliary objects A DRL. in Group object groups a number of objects and has the attribute, "Member Group of objects rather than a single object. which Relations," other Goals through a Is-a-Subgoal-of tells link. how us the objects are related. For example, a Goal The other objects in may DRL A relation can take a be related to a Group of such as Question, Procedure, and Viewpoint represent somewhat auxiliary aspects of decision making such as the questions raised and the procedures used for answering the questions. We more discuss details of DRL [Lee 1990a]. DRL al. has been implemented in a system called SIBYL, which runs on top of Object Lens [Lai et 1988]. Although the above description of provided by SIBYL for using DRL is DRL may seem quite simple decision tasks such as designing a workplace layout. create objects like a Decision Problem, Goals, menus and template are specified, cell of object. editors. Once SIBYL displays them the matrix, the user gets the a Decision Problem of all and some of shown such as creating a new Goals its in Fig. 11. it group easy to and Alternatives By mouse-clicking on a the actions that can be performed For example, by mouse-clicking on a Goal, the user can get a other Goals. for real For example, SIBYL makes and Alternatives by providing context-sensitive in a matrix such as menu complex, the actual user interface and SIBYL has been used menu subgoal, or displaying a goal tree showing on the selected containing action items how this Goal is related to in Comparative Analysis of Design Rationale Representations Close Show Goal Lalike 34 Comparative Analysis of Design Rationale Representations provides for using DRL described in more detail in [Lee 1990b]. Using the decision rationale is represented in DRL, the computer can provide among claims, propagating many and merging the al. We providing multiple discuss these services in 1988]. Evaluation of 5.3. managing the dependencies services such as plausibilities automatically, viewpoints, and retrieving useful knowledge from past decisions. [Lai et 35 DRL as a Design Rationale Language The Argument Space An argument people might which is represented in call facts, DRL sometimes arbitrary and is A as a set of related Claims. assumptions, statements, or difficult "Plausibility," that indicates how much advantage of not imposing a set of to make, confidence Claim subsumes what other Instead of rules. we have if we need to make these distinctions, has the attribute, in the claim. This has the predetermined categories on the user, and avoiding the ambiguity resulting from the disagreement among people on what Also, making DRL Claim a the distinction, say, between simply or by specializing a claim or by using nominal categories as values for the "Plausibility" attribute in different Claims. facts or assumptions We like "facts" and "assumptions" can do so post in fact facto or dynamically by using a numeric measure as the plausibility value, and mapping between measure and the measure based on the norrunal categories such when discuss different plausibility measures A are. and assumptions, we can do so facts as facts or assumptions. this We discussing the evaluation space. Claim can be Supported, Denied, or Presupposed by another Claim. These relations among the Claims allow us to respond to a Claim directly without, as in IBIS, having to indirectly respond to the Position that responds commands can be dimmed," current window would be shown and Claim for the Alternative the logical and the dynamic relation, we are making Of, is a Claim, commands distracting," rather than structure of the For example, argument more when we the claim that Claim from an Alternative achieves the goal. second Claim. For example, the Claim, "Irrelevant having not applicable to the to be formulated as a These direct relations among the Claims allow us in question. special types of Claims. relation to the directly denies the Claim "The link a 1 Claim 1 easily. to All DRL to see relations are Claim 2 through a Supports supports Claim 2. Likewise, an Achieves object to a Coal object represents the claim that the alternative Hence, any DRL relation, like Supports, Denies, Achieves, Is-A-Subgoal- and can be argued about; i.e. people can support, deny, or qualify them. For Comparative Analysis of Design Rationale Representations example, "Commands not applicable to the current window would be shown and denied by "Irrelevant commands can be dimmed". That the itself is a relational by their Claim, which is 36 then denied by first Claim is 'Dimmed commands are Unlike Toulmin's argument structure, the roles of a Claim are represented by its Datum types such as to equivalent of the Toulmin schema. Qualifier is still distracting its relations to is not shown because it is ambiguous what Figure 12 shows the or to Warrant. represented as the value of the is "Plausibility" attribute of the Supports relational claim Rebuttal " mere presence." other Claims, not by changing DRL distracting. denied by the second between the two Claims a rebuttal is at the top. denying. This representation of roles in terms of relations allows us to see the global picture of the argument structure without being confined to the micro-structure defined by the Toulmin schema and to expand the structure where and when we need deny rather than having to We can also to. deny exactly and directly the thing that ambiguously overload the meaning of we want to rebuttal. Supports Supports / Supports Claim Figure 12. The Toulmin model The Criteria Space DRL represents the criteria space fairly well. In DRL, in DRL. criteria are we corresponding goal, namely the goal of achieving the criterion, and because the richer relationship among A can always define a we want to convey these goals than what the term criteria usually conveys. example, a Coal Is-a-Subgoal-of another Coal achievement of the second. set of if achieving the subgoals can be related first Goal among themselves by creating a Group object For facilitates in various they can be mutually exclusive, independent of each other, or partially overlapping. relationships are represented DRL represented by Goals. uses the term "Goal" rather than "Criterion" because for each criterion, and specifying these Goals to the ways; These be its Comparative Analysis of Design Rationale Representations members; the the Group, as relations shown among these Goals are specified in the "Member 37 Relations" property of in Fig. 13. is-a-subgoal-of Members: Member Relations!" conjunctive partially specifying Figure 13. A Gro»p of Subgoals in DRL Decision Problem represents the goal of choosing the best alternative. the decision it means to problem are subgoals of the decision problem in the choose the best alternative. For example, the Goal "Easy Decision Problem of our example the property 'Easy to use."' if we In other interpret it to use" is a mean "Choose to words, satisfying All the other goals for sense that they elaborate what subgoal of the the alternative that has this goal facilitates the achievement of the goal of choosing the best alternative.' '^ Since the Is-a-Subgoal-of relation whether about a goal is this relational claim. should be a goal at going to is a Claim, as desirable or whether For example, In Fig. 10, there all. it is is we small implementation is relation, we can argue about it is "It should not be a goal because Mac a subgoal of the decision problem, is i.e., not a desirable property that should be used to compare alternatives. of these Claims Elsewhere [Lai DRL can argue about whether small implementation an argument, have large memory soon," denying that The record any other contributes to achieving another goal by arguing and et al. 1988], create multiple viewpoints, their relations represents the we and discuss extract how argument space this representation of Goals in for the goal space. DRL from past decisions knowledge useful allows us to to the current decision. of the model underlying DRL is more complicated and fully explained in [Lee, Roughly, for a given decision problem of the form, what is the best alternative for X, its underlying interpretation is 'the goal of choosing the best alternative for X". hs subgoal of the form, G, is strictly speaking "the goal of choosing the alternative that satisfy G". An alternative of the form, A (e.g. the global menubar at top of the screen), is to be interpreted as "Choosing the alternative A". It is in this sense that a decision problem is the parent of the other goals and that an alternative achieves a goal. This nicety, though important for computational purposes, can be ignored by The precise semantics 1991a]. human users. Comparative Analysis of Design Rationale Representations 38 The Alternative Space DRL "Commands that DRL represents only parts of the alternative space well. the specialization relation among them through at the top of the However, design window". through the specialization window" alternatives An relation. can represent alternatives and Thus, v^e the Is-a-Kind-of relation. a special case of the alternative, is may be much more complex ways related in space. These relations are more complex way, DRL example, of represents the arguments about the alternative space the alternative at alternative, all We is or whether an alternative an alternative only if ""the is commands same way at the top of the represents the it The version of another really a specialized relations, relations can be also qualified For example, window system main window" by linking the two claims with object to this Claim, in turn, design DRL. can argue about whether an alternative should be an another Claim via a Presupposes relation. window" than Next-Version-of. in the context of a by creating Claims that deny or support the appropriate Good-Alternative-for or Is-a-Kind-of. to the for beyond the current expressive power arguments about the goal space. say Alternative can be related to another Alternative via, example, the following relations: Elaborates, Simplifies, or Is-the for Alternatives can be related in a Ccin "On each we such as Is-a- by linking them to can say that "commands on each allows the attachment of a Presupposes relation. menu windows One can of course by pointing out another way of implementing the window window. The Evaluation Space In DRL, each Claim has the "Evaluation" attribute "Plausibility" The (how tells us how likely the claim overall evaluation of important the claim is true) an alternative is is, and "Degree" and (to its value is The a function of both what extent the claim is true). represented as the "Evaluation" attribute value of the relational claim, Is-a-Good-Alternative-for, Problem. and "Degree." attributes: "Evaluation," "Plausibility," between the Alternative and the Decision This value represents the extent to which the alternative satisfies the overall goal. This value, in turn, is a function of the evaluations of the Achieves relations that link the Alternative to the subgoals of the Decision Problem satisfies the subgoals). It is goal, such as the extent to also a function of how (i.e., the extent to which the alternative the subgoals interact to satisfy the parent which tradeoffs and synergies exist among these goals. Comparative Analysis of Design Rationale Representations DRL does not commit come with evaluation measures should produce evaluation measures the measure of In the global over the subgoal relations global we menubar should also know how if and merging them we want to ' ... subgoals are related menubar the probability of the achieves the goal 'Reduce screen clutter'" - combine ...' achieves the goal 'Easy to use'," given that among themselves and an algorithm based on Bayes' theorem, However, as discussed for to the That parent goal. algorithm the algorithm is, if We we also might 'In know how try to the these work out such example. in Section 2, the exact user can trust the algorithm. two Claims - commands'" and produce the probability of the Claim "The alternative to to use probability as the global menubar...' achieves the goal 'Can easily access 'In "The alternative For example, can use nominal However, such for evaluation. the algorithm for propagating at a higher level. plausibility, then "the alternative DRL measure of evaluation. The user of to a particular numeric measures, or whatever they devise categories, 39 is is important only based on to the extent that the many assumptions that the user feels are seriously violated, then the exactness of the algorithm does not contribute much. We might as well concentrate on how this philosophy and to tries to DRL takes explicit the support people for making these judgments. help by modularizing and helping to make relationship that need to be considered for these judgments. The Issue Space In DRL, the unit of the issue space Issue of gIBIS and a Question Decision Problem if Problem is, e.g. tiling or the if Is we a part a decision problem. overlapping. decision problem first A QOC. A is is A require deciding Decision Problem IS-A-KIND-OF another Decision a special case of "Where to place the "Where We are sure that there are many Also, the Is-Suggested-By relation in IBIS relation to represent on the philosophy any is other possible relations. method for may not be useful in a Is a kind of and For example, the dynamic aspect of the issue space. vague but nevertheless useful as a not described by others. However, again, that the vocabulary should better to provide a tiiat relation that is to place the window commands?" Of can relate the decision problems through the generic relations such as of. constructs For example, what the window layout a special case of the second: e.g., Replaces relation in IBIS seems important for describing the it is Decision Problem corresponds to an Decision Problem Is-A-SubDecision-Of another window commands might emacs window commands?" course, is a decision for the first requires a decision for the second. deciding where to place the algorithm in be extended to tailor the task in catch-all DRL is based hand and that extending the vocabulary as needed rather than provide general. Comparative Analysis of Design Rationale Representations 40 Relation to Other Studies: 5.4. In this section, we relate DRL to the existing representations that we discussed in the previous section. 15 The DRL structure helps us to & Rosson artifact in the sense of [Carroll its user with the designer's intention or reasons for 1990], the designer's claims captured in the design process are To often about the psychological consequences of the artifact. that extent, designer's claims explicit help us to identify the psychological design rationale. we & be careful not to confuse the claims about the psychological consequences of an artifact for the artifact [Carroll embodied by an extract the claims Although we have Kellogg 1989]. can extract from the example in Fig. 10 the following claims embodied by making the For example, the artifact of the Macintosh menubar: • The • Placing the irrelevant commands in the global window commands menubar in the global reduce the screen clutter by not duplicating the • Longer mouse travel required is not distracting menubar facilitates commands for if dimmed. ease of use because each by placing the window commands it would window in the global menu is not a serious detraction from the easy usability. Some of these claims may be obvious, but even then they are worth stating explicitly because they can then become subject to tests or further elaboration. decontextualize claims, might all global menu agree that "Longer is mouse travel required embodied by an Furthermore, not all some DRL and artifact, the claims that claims in Carroll's sense, identify is at most small. DRL but only if try to in the the screen size does not allow us is to identify all only those that come up explicitly in the course of design. come up in the course of design are worth stating as the and can be viewed as an Position to Alternative, DRL through SIBYL help tool for extracting these claims. gIBIS have similar structures at least at a high level. DRL, " the representation of design rationale in of these claims, Decision Problem in when we by placing the window commands not a serious detraction from the easy usability small so that the extra distance to travel the claims Especially, some obvious claims may not be obvious any more. For example, we and Argument Issue in gIBIS corresponds to to Claim. However, gIBIS has ArgNoter [Stefik et al. 87] is one of the earliest attempts to capture design rationale. Although DRL is designed to support many of the goals underlying Argnoter, we do not discuss ArgNoter here because no specific representation has been proposed for realizing its goals. Comparative Analysis of Design Rationale Representations quite limited expressiveness, as we argument; we have discussed above: We cannot represent or argue about the evaluation the relationship among We alternatives or arguments. Supports B without disagreeing with A nor DRL evaluating the alternatives remain implicit. ways: the explicit representation of the argument space, provision of the argument space not only cannot directly respond criteria used. cannot say that Most notably, B. 41 to an We cannot indicate we disagree that A the criteria used in the arguments can be viewed as extending gIBIS in several criteria space, the richer representation of the infra-structure for defining evaluation measures, and the for the evaluation but also for the alternative and the criteria spaces as well The gIBIS structure has the advantage of being simple, buys or costs this simplicity us. represent most of the discussion, although it among If components the use the language, then used at However, all. of design rationale. and B but not they want to A that objects to B. say and yet From a computational with them. Also, we this is If natural It is DRL what they want for people more expressive should remember that often in the context of structure. But QOC perhaps the closest is SIBYL not any is an empirical question, and it is DRL to to say, e.g., that to use, we would structures it is not only the language more cumbersome than we plan and Coal subgoals maps to A what do more to itself we things but also the believe that the the use of the gIBIS less clear to us. Decision map roughly to Criterion. Claims Mini-Theory depending on whether the claims are empirical statements or , parts of an established theory, or an informal theory. how agree with Both have the five spaces Problem maps not clear I to find out. at least in its basic structure. to Question, Alternative to Option, be prefer such a language. would allow us argument space are is to People can be usability. clearly delineated, although the constructs for the to Data, Theory, or enough of the relations better to be simple than not to always enhance user interface that determines usability. Based on our limited experiences, use of general is many a language with a finer structure allows people to say enough point of view, an empirical question what Argument) simple structure makes people more likely to a simpler structure does not frustrated with not being able to say exactly is it does not explicitly reveal worth the while. certainly it is and structure {Issue, Position, Its argue about these to the QOC concept of Bridging However, we pointed out constructs for the argument space. Criteria. Also, at least some earlier that The concept it of of the links can be argued about: e.g. the evaluation link between an option and a criterion can be supported by Data or Theory. WTiether the relation we can argue about other between Data and Theory, is links, for not clear. example, the relation among Criteria or Comparative Analysis of Design Rationale Representations It seems that QOC That language. is is, currently more an attempt is it a to model rather than 42 a fully-developed representation understand and categorize the elements of design rationale without providing yet a specific set of vocabulary for expressing them. a worthy endeavor and consistent with their warning against representation. Given the similarity in the underlying structure, of course It is premature commitment a we hope DRL that representation language adequate for representing most of the elements that the to a provides a QOC research has been articulating. many There are Answers in between PHI and DRL. parallels PHI corresponds among quasi-hierarchical relation The quasi-hierarchical relation Denies relations in DRL. But Issues in PHI corresponds among Arguments many constructs in Goals, Presupposes, or Is-a-Subgoal-Of, further the extensions that The quasi-hierarchical to the Is-a-Kind-Of relation in the alternative PHI made complex relations and making PHI in DRL and Is-a-Part-Of to IBIS explicit is among DRL. The in to the Is-a-Subdecision-of relation. and specialized into the Supports no correspondence find . relation space DRL Thus, in PHI, can be viewed as pushing by generalizing the hierarchical structure some other elements, such as to more especially those in the criteria space. & In [Potts Bruns 1988], IBIS representation. used as the representation for the rarionale component is The component modular enough, however, is alternate representation for the module. [Fischer et VIEWPOINTS. Again, DRL can be used which we believe is more is more expressive and can 6. CONCLUSIONS A large body can get and how The easily we is its issue base, CRACK, because DRL better support icnowledge-base operations. task of using choice of representation in their can be viewed as an 1989] also uses IBIS for natural to interface with the other component, of research in the last 1985]. DRL as the alternative representation for the module, two decades or so points right representation for a given task [Amarel 1968; Levesque al. that Bobrow to the 1975; and reusing design rationale is importance of choosing a Winston 1984; Brachman no different. The can get them depends heavily on the representation especially important as in design rationale capture, because a task altogether, attributing the failure when wrong a human is we use. & we The the user of the representation, representation can turn people and frustration benefits away from to the task itself rather the than the inadequacy of the representation used. People might conclude that capturing design rationale is not worth the trouble because it is so hard and because it does not provide enough rewards for 43 Comparative Analysis of Design Rationale Representations the efforts. But the real problem might be that the representation does not allow us to represent easily what we want way to represent or in a we know how important that to much that can provide Thus, benefit. evaluate a representation for a given task, is it our case, in for capturing design rationale. In this paper, we made a step forward by characterizing the domain of design rationale, by i.e., identifying the kinds of elements that form the rationale as well as the relations that hold Characterizing this domain among them. represent, what we have decided not helps us to map is we important because to represent, and what are know what we can then their consequences. different parts or aggregates of the domain. In other words, it We Using the defined the scope of the existing representations and discussed their adequacy. we have also presented a language, called DRL, which we believe most of the existing languages and overcomes many natural to human However, users. that is a testable claim that by many DRL DRL, which we hope will be the topics for future research with The step we made, however, many is understand the important issues We a small one and in to represent the its We also users. we we have a long for way to tractability. We We also need to so that need much for the task other categories: many we have human to articulate the characteristics that identify the computational services that we know what us. go before more we fully in make we usability tfian being able other characteristics, provide abstractions, that are important [Bobrow 1975]. we mentioned two is still discussed the limitations of We need to think hand, and in what they matter. Then, there are other dimensions to a representation than In Section 2, that evaluating a design rationale representation along elements in the domain explicitly or not. There are ability to way plan to investigate by others as well as by expressive power. Even then, expressiveness involves about whether these characteristics matter way more expressive than designing an ideal representation for representing design provided a framework one dimension - such as the tasks is of their limitations in a empirically by using rationale. also provides a framework for defining the scope and assessing the expressive adequacy of a representation. framework, It them with the the different meanings of design rationale by associating its expressiveness. and computational a representation more usable. can provide with design rationale they require and any tradeoff between their requirements and other requirements such as hunnan usability. There are some existing studies which address these problems [Lee 1989; Yakemovic Newman & 1991; Fischer et al. Marshall 1990; Yakemovic 1991; Lee 1991b; focus on the representation being used, and Shum 1991]. We & Conklin 1990; Conkiin & need more such studies, more more systematic categorization of the results. We Comparative Analysis of Design Rationale Representations believe that the benefits from explicit representation of design rationale for the efforts that we 44 would more than pay put into such studies. Acknoivledgments. The whole project of articulating the domain arose from our frustration with not knowing exactly to Tom Malone DRL. how to evaluate the existing representations suggested that we which components of a decision matrix One analysis here. great environment in Jeff Tom Conklin, influenced which he wrote much by the Mark MIT AI This to relate them it makes like to That insight triggered much of the explicit. thank Frank Halasz for providing him with the this paper. The comments from anonymous the reviewers, Moran, Randy Trigg, and Austin Henderson were valuable. The paper was the Danny Bobrow, Frank Rogers, would of us (Jintae) and how define the scope of a representation by thinking about Stefik, comments from Halasz, Bill and Norbert the members ARG of (Argumentation Reading Group): Janssen, Cathy Marshal, Susan Streitz. We also Lab, especially Patrick Winston, Rick Lathrop, and work was supported, in part, Newman, Dan Russell, Russ thank the members of the learning group at Gary Borchardt. by Digital Equipment Corporation, the National Science Foundation (Grant Nos. IRI-8805798 and IRI-8903034), and DARPA (Contract No. N00014-85-K- 0124). REFERENCES Amarel, & On S. (1968). the representation of Problems of Reasoning about Actions, in Webber, B. N. Nilsson (Eds.) Readings Alto, in Artificial Intelligence. Tioga Publishing Company: Palo CA. Bimbaum, M. Flowers L., & R. McGuire Towards an AI Model (1980). Bobrow, D. (1975). Dimensions of Representation, Representation and Francisco, Bobrow, D. Understanding: Studies in in Bobrow, D. Cognitive Science. of Argumentation. Stanford, Proceedings of the First National Conference on Artificial Intelligence & CA. A. Collins (Eds.) Academic Press: San CA. & Intelligence Goldstein (1980). Representing Design Alternatives. Proceedings of and Simulation of Behavior Amsterdam, Netherland. I. Artificial R. & H. Levesque (1984). The Tractability of Subsumption in Frame-Based Description Languages. Proceedings ofAAAI-84 pp. 34-37. Austin, TX. Brachman, Brachman, R. Kaufmann Card, S., ]. & H. Levesque, Eds. (1985). Readings Los Altos, CA. Publishers, Inc. D. Mackinlay Proceedings of CHI '90 & in Knowledge Representation. Morgan : G. G. Robertson (1990). 117-124. Seattle, WA. The Design Space of Input Devices. Comparative Analysis of Design Rationale Representations 45 Human Computer Interaction Scenarios as a Design J. & M. B. Rosson (1990). Representation. Proceedings of Proc. of HICSS-23: Hawaii International Conf on System Sciences. Los Alamitos, CA. IEEE Computer Society Press. Carroll, Carroll, & J. M. Rosson B. Deliberated Evolution: Stalking the (1991). Space. Human-Computer Interaction Carroll, M. J. & R. L. Human-Computer Campbell View Matcher in Design (this issue). The Case of Artifacts as Psychological Theories: (1989). Interaction. Behaviour and Infonnation Technology 8(4) 247-256. Artifact as Theory-Nexus: Hermeneutics Meets TheoryJ. M. & W. A. Kellogg (1989). Based Design. Proceedings of CHI '89 7-14. Austin, TX. Carroll, Conklin, J. ACM Conklin, & M. L. Begeman (1988). gIBIS: A Hypertext Tool Transactions on Office Information Systems & J. K. C. Yakemovic Hutrmn-Computer Lemke, R. McCall Human-Computer Interaction Fischer, G., A. Fischer, G., R. McCall & A. & H. Rittel (1970). & ACM Malone & K.-C. Exploratory Policy Discussion. Process-oriented Paradigm for Design Rationale. Morch A. (1991). Making Argumentation Serve Design. (this issue). Morch Design Environments (1989). of CHI' 89 for Constructive Yu of California, Berkeley, (1988). and pp. 269-276. Austin, TX. Center for Planning Issues as Elements of Information Systems. and Development Research Univ. Lai, K.-Y., T. for pp. 303-331. (this issue). Argumentative Design. Proceedings Kunz, W. A (1991). Interaction 6(4) Object Lens: Transactions on Office Information Systems A 6(4) Working Paper. "Spreadsheet" for Cooperative Work. 332-353. Task Embedded Knowledge Acquisition through a Task-Specific Language. J. Detroit, MI. Proceedings of IJCAI Workshop on Knowledge Acquisition Lee, (1989). Lee, J. (1990a). SIBYL: Shellard (Eds.) Cambridge, A Qualitative Decision Artificial Intelligence at Management System, MIT: Expanding in Winston, Frontiers, vol. 1. P. H. MIT & S. Press: MA. SIBYL: A Tool for Managing Group Decision Rationale. J. (1990b). Computer Supported Cooperative Work LA, CA. ACM Press. Lee, Proceedings of Lee, Decision Rationale Management System: Computational Representation and J. (1991a). Use of the Reasons for Decisions. Ph.D. Thesis forthcoming. Department of Electrical Engineering and Computer Science, MIT. Lee, J. (1991b). Extending the Potts and Bruns model for recording design rationale. Proceedings Austin, TX. of 13th International Conference on Software Engineering Lewis, Problem-Centered Design for Expressiveness and Facility J. Rieman & B. Bell (1991). Graphical Programming System. Human-Computer Interaction (this issue). C, in a Lowe, D. (1986). Information. SYNVIEW: The Design of a System for Proceedings of Computer Supported Cooperative Cooperative Structuring of Austin, TX. Work Comparative Analysis of Design Rationale Representations MacLean, A., R. Young, V. Bellotti Elements of a Design Rationale & 46 Moran (1991). Questions, Options, and Criteria: User Interfaces. Human-Computer Interaction (this T. for issue). MacLean, & M. Young A., R. Artifact. T. P. Moran Proceedings of CHI' 89 (1989). Design Rationale: The Argument behind the Austin, TX. Exploring Representation Problems Using Hypertext. Marshall, C. (1987). Hypertext '87 Proceedings of PHIBIS: Procedurally Hierarchical Issue-Based Information Systems. McCall, R. (1987). Proceedings of Conference on Planning and Design in Architecture Society of Mechanical Engineers. Boston, McCall, R. (1990). PHIDIAS: A PHI-based Design Environment integrating Dynamic Hypertext. Proceedings of European Conference on Hyupertext MA. American CAD Graphics into Versaille, France. A Semantic Analysis of the Design Space of Input J., S. Card & G. Robertson (1990). Devices. Human-Computer Interaction 5(2-3). McKinlay, Mostow, Newman, Toward (1985). J. S. & Better C. Marshall & /4/ Mflgflzme 6(1) pp. 44-57. Pushing Toulmin Too Far: Learning From an Argument Xerox PARC, Technical Report. (1990). Representation Scheme. Potts, C. Models of the Design Process. G. Bruns (1988). Recording the Reasons for Design Decisions. Proceedings of 10th pp. 418 -427. International Conference on Software Engineering Shum, S. (1991). to Cognitive Dimensions of Design Rationale, submitted to HCr91 Lewis (1988). Computer-Aided Reasoned Discourse or. How Argue with a Computer, in Guindon, R. (Eds.) Cognitive Science and Its Application to Smolensky, P., B. Fox, R. King Human-Computer Streitz, N., ]. & C. Interaction, pp. 109-162. Ablex: Hannemann & M. Thuming (1989). Travelling through Activity Spaces. Toulmin, Winston, S. (1958). P. (1984). The Uses of Argument. Artificial Intelligence. From Norwood, NJ Ideas and . Arguments Proceedings of Hypertext '89 Cambridge Univ. Press: to Hyperdocuments: pp. 343-364. Cambridge, England. Prentice Hall: Reading, MA. Yakemovic, K. C. B. & J. Conklin (1990). Observations on a Commercial Use of an Issue-Based LA, CA. Information System. Proceedings of Computer Supported Cooperative Work il JJie International Center for Research on the Management of Technology Sloan School of Management Massachusetts Institute of Technology Working Paper List Number Date 1-89 11/89 Author(s) Title Netgraphs: A Graphic Representation of Adjacency Tool for Matrices as a Network Analysis 2-90 Strategic Transformation 3-90 Managing CAD Systems and the Success in of High Technology Companies The Personality and Motivations Roberts Robertson Allen Mechanical Design Engineering 1/90 (Rev. 3/91) 4-90 George Allen of Technological Entrepreneurs 1/90 5-90 Current Status and Future of Structural Panels in the Wood Products Industry 6-90 Do Nominated Boundary Spanners Become Effective Technological Gatekeepers? 7/90 -90 The Treble Ladder Revisited: as They Grow Older? Why Do Engineers Lose Interest in the Dual Ladder 8/90 10-90 The Emergence of Fiber Optics Professionals: A People and Technology Transfer Exploring the Dynamics of Dual Ladders: A Longitudinal Study 8/90 12-90 8/90 13-90 8/90 14-90 Managing the Introduction in a Multi-Plant of New Process Technology: International Differences Networic Task Characteristics and Organizational Problem Solving Process Change The Impact of "Sticky Data" in Technological on Innovation and Problem-Solving 8/90 15-90 5/90 16-90 7/90 McCormack Utterback Organizational Relationships and Advancement of Technical Ten Year Longitudinal Study in One Organization Work Environment, 8/90 11-90 Allen Katz Technological Discontinuities: 8/90 9-90 Allen Nochur 6/90 (Rev. 7/91) 7-90 Montrey Utterback 4/90 Underinvestment and Incompetence as Responses to Radical Irmovation: Evidence from the Photolithographic Aligriment Equipment Industry Patterns of Communication Among Marketing, Engineering and Manufacturing — A Comparison Between Two New Product Teams Working Paper List (continued) Number Date 17-90 Author(s) Title Age, Education and the Technical Ladder Allen 9/90 (Rev. 8/91) 18-90 Katz A Model of Cooperative R&D Among Competitors 1/90 19-90 4/90 20-90 Strategy, Structure, the and Performance in Product Development: Observations from Auto Industry Sinha Cusumano Cusumano Nobeoka Organizing the Tasks in Complex Design Projects Eppinger Whitney Smith 6/90 Gebala 21-90 The Emergence of a New Supercomputer Architecture 7/90 22-90 Superceded by 39-91. 23-90 Software Complexity and Software Maintenance Costs 8/90 24-90 Leadership Style and Incentives Banker Datar Kemerer Zweig Rotemberg 9/90 25-90 Afuah Utterback Saloner Factory Concepts and Practices in Software Development Cusumano 11/90 26-90 Going Public: Sell the Sizzle or the Steak Roberts 10/90 27-90 11/90 28-90 Evolving Toward Product and Market -Orientation: The Early Years of Technology-Based Firms The Technological Base of the New Enterprise Roberts Roberts 11/90 29-90 Innovation, Competition, and Industry Structure 12/90 (Rev. 6/91) 30-91 Product Strategy and Corporate Success Roberts 1/91 31-91 Utterback Suarez Meyer Cognitive Complexity and CAD Systems: Beyond the Drafting Board Metaphor 1/91 Robertson Ulrich Filerman 32-91 CAD System Use and Engineering Performcmce in Mechanical Design 1/91 33-91 6/91 34-91 5/91 Robertson Allen Investigating the Effectiveness of Technology-Based Alliances: Patterns and Consequences of Inter-Firm Cooperation The High Centrality of Boundary Spanners: Organizational Networks A Source of Natural Efficiency in George Allen Ancona George Robertson Working Paper List (continued) Number Date 35-91 2/91 Title Impacts of Supervisory Promotion and Social Location on Subordinate Promotion RD&E Setting: An Investigation of Dual Ladders in an Allen (Rev. 11/91) 36-91 Demography and Design: Predictors of New Product Team Performance 1/91 37-91 Author(s) Katz Tushman Tlie Changing Role of Teams Ancona Caldwell in Organizations: Strategies for Survival 2/91 38-91 Schrader Informal Alliances: Information Trading Between Firms 3/91 39-91 3/91 40-91 3/91 41-91 Supplier Relations and Management: and U.S. A Survey of Japanese, Japanese-Transplant, Maneuvering and Mass-Market Dynamics: The Triumph Strategic of VHS Over Beta The Software Cusumano Takeishi Auto Plants Factory: An Entry for the Encyclopedia of Software Engineering Cusumano Mylonadis Rosenbloom Cusumano 3/91 42-91 Suarez Utterback Dominant Designs and the Survival of Firms 4/91 (Rev. 7/92) 43-91 An Environment for Entrepreneurs Roberts 6/91 44-91 7/91 45-91 Technology Transfer from Corporate Research of Perceptions on Technology Adoption When Speeding Concepts to to Operations: Effects Nochur Allen Market Can Be a Mistake Utterback Meyer 3/91 Tuff Richardson 46-91 Paradigm Shift: From Mass Production to Pine Mass Customization 6/91 47-91 8/91 48-91 The Marketing and 10/91 (Rev. 2/92) 49-91 Murotake Computer Aided Engineering and Project Performance: Managing a Double-Edged Sword R&D Allen Interface Systematic' Versus Accidental' Reuse in Japanese Software Factories 10/91 50-91 Flexibility and Performance: A Literature Critique and Strategic Framework 11/91 51-91 11/91 52-91 12/91 53-91 12/91 Shifting Econonues: From Craft Production to Flexible Systems and Software Factories Beyond Understanding the Commitment of Pioneers in Emerging Fields and Technology Persistence: of Science Institutional Variations in Pioneering Researchers Problem Choice and Persistence among Cusumano Working Paper List (continued) Number Date 54-91 Title The Role of Students in Pioneering New Fields of Science and Technology 12/91 55-91 Debackere Rappa Technological Communities and the Diffusion of Knowledge 12/91 56-91 Debackere The Voice of the Customer Griffin 10/91 57-92 Hauser The Influence of Inter-Project Strategy on Market Performance in the Auto Industry 1/92 58-92 1/92 59-92 Nobeoka Cusumano Linking International Technology Transfer with Strategy and Management: A Literature Commentary Using the Literature in the Cusumano Elenkov Study of Emerging Fields of Science and Technology Rappa Garud 7/91 60-92 Author(s) Rappa Technological Progress and the EHu'ation of Contribution Spans 12/91 Rappa Debackere Garud 61-92 9/91 62-92 Technological Trajectories and Selection Mechanisms in the Development of Cochlear Implants On the Persistence of Researchers in Technological Development An International Comparison of Scientists in an Emerging Field Life on the The Social Construction of Technological Reality Frontier: Core Competencies, Product Families and Sustained Business Success 2/92 66-92 4/92 68-92 2/92 Creating Occasions For Technological — - Bennett the Renaissance™ Spirometry System: Listerung Voice of the Customer Puritan to the Time Flies When You're Having Fun: How Consumers Allocate Their Time When Evaluating Products (Rev. 11/92) 69-92 Moving Ideas to Market and Corporate Renewal 7/92 70-92 Utterback Meyer Windows Of Opportunity: 3/92 Adaption In Organizations (Rev. 9/92) 67-92 Garud Rappa 1/92 65-92 Debackere Rappa 1/92 64-92 Garud Rappa 12/91 63-92 Garud Rappa Tyre Orlikowski Hauser Hauser Urban Weinberg Meyer Utterback Project Management in Technology Innovation, Application and Transfer Frankel 5/92 71-92 Investments of Uncertain Cost Pindyck 9/92 72-92 Identifying Controlling Features of Engineering Design Iteration 10/92 73-92 10/92 Smith Eppinger Objectives and Context of Software Measurement, Analysis and Control Cusumano Working Paper List (continued) Number Date 74-92 Tide An Empirical Study Author(s) of Manufacturing Flexibility in Printed-Circuit Board Assembly Suarez Cusumano 11/92 Fine 75-92 11/92 76-92 Japanese Technology Management: Innovations, Transferability, and the Limitations of 'Lean' Production Hauser Customer-Satisfaction Based Incentive Systems Simester Wernerfelt 11/92 77-92 The Product Family and the Dynamics of Meyer Core Capability Utterback 11/92 78-92 11/92 79-92 Multi-Project Strategy and Organizational Coordination Nobeoka Cusumano in Automobile Product Development Pattern of Industrial Evolution: Dominant Design and the Survival of Firms 12/92 80-92 11/92 81-92 Cusumano King Innovation from Differentiation: Pollution Control Departments and Iimovation in the Printed Circuit Industry Answer Garden and Utterback Suarez Ackerman the Organization of Expertise 1/92 82-92 Skip and Scan: Qeaning Up Telephone Resnick Virzi Interfaces 2/92 83-92 Developing New Products and Services by Listening to the Voice of the Customer Roberts 8/92 84-92 A Comparative Analysis of Design Rationale Representations 3/92 85-93 1/93 An Introductory Note for Using Analyze Nodes, Relationships, Partitions and Boundaries Relational Data in Organizational Settings: AGNI and Netgraphs to Lee Lai George Allen The International Center for Research on the Maiugement of Technology Working Paper Order Form Name: Title: Company : Address: n I would like to become a working paper subscriber and during the current year. ($125 D I would like to U.S., Canada, Mexico; $150 all # # papers published other countries) order working papers individually. Please send papers: # receive all me the following ^)668 Date Due MAY 2 5 2002 .IBRARIES DUPl 3 TOAD OOflSMDflT M