A decision rationale management system : capturing, reusing, and managing the reasons for decisions

Texte intégral

(1)

(2)

(3)

(4)

(5) DEWEY HD28 .M414. A. Decision Rationale. Management System:. Capturing, Reusing, and Managing the Reasons for Decisions Jintae. Sloan School. Lee. WP #3505-92. CCS TR #136 December, 1992. Submitted to the EECS Depanment on September 17, 1991, in partial fulfillment of the requirements for the degree of Doctor of Philosopy in Computer Science..

(6)

(7) A. Management System: Capturing, Reusing, and Managing the Reasons for Decision Rationale. Decisions. by Jintae. Submitted. to the. EECS Department on September. in partial fulfillment. for the degree of. Lee. 17, 1991,. of the requirements. Doctor of Philosophy. in. Computer Science. Abstract. This thesis identifies the needs for capturing and managing decision rationales, aniculates the concept of a decision rationale. management system. that. meet these needs, and presents. a computer system that implements the concept-. Capturing and managing decision rationales, bring about. many. benefits.. The. i.e.. the deliberations leading to decisions, can. rationales can then be used to. suppon decision making,. can be shared among decision makers, and can be reused for similar decisions. rationale. management system,. i.e.. a system that captures and. to provide these benefits, requires three. A. manages decision. decision. rationales. major components: a language for representing. elements of rationales, a method of using the language to capture the rationales, and a services that use the captured rationales to. of. suppon decision making.. This thesis articulates a model of decision rationales and uses for representing the elements in this model.. system. set. The. that helps people to capture rationales in. it. to. develop. thesis also presents. DRL. DRL,. SIBYL,. a language a. computer. by providing a number of interface. features intended to reduce the overhead associated with explicit representations. Using the. rationales captured in. DRL, SIBYL. provides computational decision services, such as. retrieving useful rationales from past decisions, maintaining dependencies. among. various elements of rationales, and keeping track of multiple decision states.. the. These. services realize the benefits of a structured representation of rationales, and provide further. motivation for capturing rationales in. DRL.. Thesis Supervisor: Patrick H. Winston Tide: Professor.

(8)

(9) Acknowledgements. First.. thank the members of. I. W. Malone, Randall. my. thesis committee. Professors Patrick H.. Davis, and Marvin Minsky, for their guidance and patience. Patrick has. supplied the enthusiasm for this research that. with his ideas that are. now. difficult to. made. cntique, which undoubtedly. me. I. Winston, Thomas. shake. the research. needed. I. all. along.. Tom. Randy has been. off.. more. has brainwashed. me. the source of constant. painful but better.. Marvin has supplied. with his vision and pointers that only he could have given.. thank. Yew. my. all. Lai.. friends at the. AI Laboratory and. the Center for Coordination Science:. Kum-. Lukas Reucker, Rick Lathrop, Kevin Crowston, Franklyn Turbak, Paul Resnick,. John Mallery, Roger Hurwitz, Gary Borchardt, Jolene Galegher, Mark Ackermann, and. many. others.. computer. Kum-Yew. --. with his interests ranging from. interaction, finance, civil engineering,. computer-supponed cooperative work,. systems dynamics, economics to mention only a few. whom he. I. — was. could always relied on for bouncing out ideas.. MIT. left. to. become. a student of finance.. fuzziness in this thesis because. out to me.. Lukas - amid. if. I. He. a friend. who. couldn't say. else.. --. provided. hold him partially responsible for any error or. he stayed around, he would have undoubtedly pointed them. the thousand activities that he. me. with the most detailed. His comments made reading. somehow. I. comments on. this thesis less painful than. also been an invaluable source of. guess. no and. has been sorely missed ever since. juggles, ranging from. parachuting, scuba diving, organizing lab tours, lecture series committee,. mention a few. human. artificial intelligence,. could wnte a book detailing. apprecianons.. this thesis than. club, to. anybody. would be otherwise. Rick has. encouragement and ideas when. my. Anemie. I. was down and. out.. I.

(10) I. would. like to. Tom. thank "design rationalizers":. Moran, Jack Carroll. Jeff Conklin, Ray. McCall. Gerhard Fischer, Allan MacClean, Simon Shum, Mark Novic, and others helped. me. directly or indirectly see better the issues involved in. would also. Marshall, Susan Brodie, and. saved. research.. many. At. others that. many many more.. last,. talked to in these places.. who came aaoss my. despite our brief interaction,. I. Laboratory: Frank Halasz, Cathy. Working. at these places not. path and. comes. to. my. way. work out and. to. left their. mind, but. only. I. marks. am. in this. sure there. apologize for not being able to thank them individually, but that. me. all,. who. I. thank. my. with smiles and. who have been. parents-in-law. PARC and GTE. I. is. running out of space and dmc.. but most of. always greeted parents,. am. I. There are others. Randy Trigg,. I. Xerox. family from starving, but also was a wonderful. ideas.. only because. at. rationales.. Newman, Austin Henderson, Danny Bobrow, Frank Manola, Michael. me and my. bounce off. are. thank people. like to. managing. who. family,. made me. very patient with. Hyokyung, Youngwon, and Jangwon, who. smile even during. my. I. thank them. all. darkest hours.. progress and always supponive.. put their daughter in the trust of a. yet never complained.. my. man who. from the deepest of. my. earns. heart.. And my. And my. minimum wages and.

(11) Contents. 1. Introduction. 1.1. 1. 2.. 3.. A. .2. 4. Summary. 9. of Approaches. 14. The Structure of Decision Rationales. 3.2. What do we want. to. 30. do with decision rationales?. 31. Models of Decision Rationale. 32. DRL: A Decision Representation Language 4.. 1. 4.2.. 5.. The Problems. Scenario. 3.1. 4.. 1. 42. Overview. DRL's Representation of. SIBYL: An Environment 5. .. for. the Decision Rationale. Using. DRL. 5.3 Setting. up an. Augmenting. 5.6. Making. 46 52. 57. Initial. Releasing the. 5.5. Model. 53. Overview. 5.2 Implementation. 5.4. 41. Structure. Initial. Structure. the Decision Structure. Outcome of. 64. 65. 75. the Decision. 5.7 Evaluating the. 58. the Decision. 76.

(12) 6.. Computational Services. 6.. Precedent Management. 6.1.1. 6.. 6.2. 1. .2. 81. Specific Retrieval Request. General Retrieval Request. Dependency Management 6.2.1. 6.3. 79. The Scope. 7.2. 8.. 1. 6.2.3 Implementation. II. Ill. 07. 113. 6.3.. Representation. 114. 6.3.2. User Interface. 116. Implementation. to Related. Systems. 120. 121. Work. that. Capture Rationales. Semi-Formal Rationale Management Systems. Conclusion. 8.. 04. 1. 6.4 Other Services. 7.1. 1. 104. Viewpoint Management. Comparison. 85. 6.2.2 Implementation. 6.3.3. 7.. 81. Contributions. 8.2 Future Research. 128. 129. 1. 34. 148. 148. 152. Appendix: Details of DRL. 157. References. 1. 77.

(13) Chapter. 1. Introduction. A. structured representation of decision rationales,. decision, can bring about. many. benefits.. i.e.. The knowledge. the deliberations leading to a. that decision. makers bring. to the. decision becomes available for other people or computational agents to share, augment, and. argue about. The representation of a decision making process serves as a document of the decision developed,. which. in turn. how. can serve as a basis for learning and justification.. In. addition, a well-structured representation provides a basis for defining computational. services for decision making, such as keeping track of dependencies and retrieving useful rationales. Some. from past decisions.. of these benefits have been explored so far by a few systems, but most of these. systems require highly formalized and structured domain knowledge. For example, the.

(14) research on derivational analogy systems [Carbonell 1986;. 1989; Steinberg. &. Huhns. & Acosta. 1988;. Mostow. Mitchell 1985] explore ways of capturing the reasoning behind design. and reusing parts of the reasoning trace for solving redesign or similar design problems. These systems, however, require that the. problem solver can run. domains where to be. the. knowledge. is. that the. domain knowledge be. in the first place.. knowledge has been formalized, decisions --. a result, they are inapplicable to. less structured, not well. formally represented. Even in the domains. knowledge. As. (e.g.. sufficiently formalized so. understood. yet, or too. design of circuits or parsers) whose. typically involve consideration of other worldly. such as availability of resources or political influences. difficult to formalize.. expensive. Nevertheless, capturing and managing rationales. is. --. which may be. no. less. important. or urgent in these domains.. The goal of. way. this thesis is to. that their benefits. This goal. is. develop a system that captures and manages rationales. such a. in. can be realized without requiring the formalized domain knowledge.. achieved by developing a representation that allows informal description of. domain knowledge. to be. included. in. formal structures. These formal structures need to. capture generic knowledge about decision making, and suppon interesting computational. Because people are. operations.. system that uses 1988], requires. this type. still. necessary to interpret the informal descriptions, any. of representation,. much human. interaction.. i.. e.. semi-formal representation [Lai. et al.. helps solve the problem of. brittle. Nevertheless,. it. performance because people can supply the expertise or commonsense where the system cannot, at least until. understand. The. last. it. we. gradually understand. more of. the. domain and make. the system. too.. point about gradual understanding leads to another goal of this research, that of. producing a system that. is. practically useful.. Apan from. its. obvious merit,. this. goal. is. derived from the more ambitious goal of producing an automated rationale management.

(15) Although informal descriptions provide us with. system.. formalized. if. one's goal. goal of automation. realizing this goal.. need. to be. automate rationale management as much as possible. This. similar to the one underlying the derivational analogy systems. is. mentioned above, but. to. is. flexibility, they. this research takes the. The approach. is. approach of incremental formalization. to start with a. system. that is useful. in. even without much. understanding of the domain knowledge, but have the system record the domain. knowledge. in the. course of. system, thereby making. it. its. that captures. it. is. is. might be. in the. into the. possible in a rationale. the rationales captured by the system. embody knowledge. form of informal descriptions. Thus, a system. and manages rationales also helps us gain more understanding of the domain. and formalize the knowledge. For system. which can then be formalized and fed back. more powerful. This feedback loop. management system because about the domain, though. use,. useful. enough. to be. this. used. approach to work, however,. in real situations.. current system cannot be claimed without. much. it. is. important that the. Although the usefulness of the. qualification, the attempt to. make. a useful. system has generated many constraints and has been achieved to a degree, as discussed. in. the thesis.. The. thesis proceeds as follows.. number of concrete problems approach that. In the rest. that. of. motivated. this chapter,. this. this thesis takes to solve these. I. first. descnbe and categorize. research (Section 1.1).. problems (Section. 1.2).. I. a. then discuss the In particular,. I. propose the concept of a decision rationale management system as a solution to these. problems and articulate In. Chapter. rationale. which. 2, a. its. components. This discussion provides a preview of the. scenario illustrates concrete behaviors that. management system should have. This scenario. the success of this thesis. is. measured.. I. thesis.. believe an ideal decision. serves as a yardstick against.

(16) The next four chapters. constitute the. main body of. elements of a decision rationale that need. models which rationales.. this thesis.. to be represented. Chapter 3 identifies the. by building a sequence of. differentiate, in different degrees, the internal structure of decision. These models are also used as a framework for. later discussions.. these models. Chapter 4 presents a rationale representation language, called. DRL. that use. the environment that. SIBYL. to capture. and. 4, 5,. 6,. SIBYL. rationale. SIBYL. space are discussed. summarizing. 1.1. its. in detail. to capture rationales.. Chapter 6. provides by using the captured rationales.. support decision making.. are located along. and the reusability of the rationales (Section in this. Chapter 5 describes. services.. to other tools thai. management systems. DRL. rationales.. describe each of the components of a rationale. management system: language, method, and. Chapter 7 compares. and manage. provides for using. discusses the computational services that. These three chapters,. DRL (Decision. Chapters 5 and 6 present SIBYL, the decision rationale. Representation Language).. management system. Based on. two dimensions:. 7.1).. Then, the tools. (Section 7.2).. First, the existing. the formalization required that are closest to. SIBYL. Chapter 8 concludes the thesis by. contributions and discussing the topics for future research.. The Problems. This research grew out of environment.. We. had. my. experience. to decide, for. in. helping a group to set up. its. computing. example, which workstation to buy, which network. protocol to use, and whether to use a relational database or an object-oriented database.. One. of the most frustrating aspect of this experience was the realization that although. hundreds of groups must have made similar decisions. in the past,. we had. to start. from.

(17) scratch because the information that decisions,. i.e.. must have been collected and analyzed. their decision rationales,. groups would make similar decisions. were not available. in the future,. difficulty in reusing rationales is only. Likewise, hundreds of. to us.. and they would not be able. from our experience because they would not have access one of the problems. to. in these. to benefit. our decision rationale. The. that. we. face in decision making.. Other problems include keeping track of the issues discussed and those yet to be resolved, explaining the rationale to other groups, or even just physically getting people together to talk. about these issues.. These problems can be categorized as follows. Consider a decision making process as consisting of sessions, e.g. meetings in case of group decision deliberations otherwise (Fig. l.I). First, there are problems oi. a session like being able. Then. to find out. there are problems of. managing rationales across sessions in. --. for. example,. previous sessions and what need yet to be. Across decisions, there are problems of sharing rationales across decisions. taking place concurrently, such as sharing information decisions.. managing rationales within. what depends on what or ask "what-if questions.. remembering what have been resolved resolved.. making or individual. among groups making. Finally, also across decisions but separated in time, there are the. reusing rationales from past decisions, as discussed above.. similar. problems of. Each of these categories. is. described in the rest of this section.. Managing Rationales within. In. a Session. making decisions, we often need. to. keep track of dependencies, compare multiple. viewpoints, and ask "what-if questions. There are the. problems. in this. category. --. for. many. techniques and tools that address. example, decision analysis and simulation. tools.. In.

(18) C a decision. ^. Managing Rationales. within a Session. Reusing Rationales across Decisions. Sharing Rationales across. Decisions.

(19) Managing Rationales across Sessions. Different kinds of problems arise in. rationales across sessions within a decision. Here, typical problems include those of bookkeepmg, for example. making process. remembering. managing. the issues discussed previously, the issues yet to be resolved, or. being done about them. Associate with this category. is. individuals involved in decision. making can exchange. manner.. making takes place. Typically, decision. necessitates finding a place. This problem, as process.. and a time. in. many people complain,. In addition, face-to-face. the. problem of ensuring. what. is. that the. their ideas in a. quick and focused. in face-to-face. meetings, which. which the decision makers can. all. meet together.. often unnecessarily prolong the decision. meetings have other known problems.. wait until the meeting to contribute their knowledge, or. people have. to. their ideas if. meetings are dominated by certain individuals. making. For example,. may. not contribute. Sharing Rationales across Decisions. Often, the different subgroups need to communicate their rationales to other groups so that the groups reach a. For example,. in. common. understanding of the problems and share their. designing a product, designers might want to. group wants some feature and dislike others. Another problem. two groups making similar decisions want I. worked. for. benefiting from shared information.. some knowledge but. systematic.. expenise.. the marketing. category arises. For example,. at. when. one point. two groups making similar decisions about hardware and software platforms.. The two groups shared many requirements and. transfer. know why. in this. to share information.. own. alternatives,. By vinue of. this transfer. hence were. being involved. in the position of. in both,. I. was. able to. was accidental and could have been more.

(20) Reusing Rationales from Past Decisions. We do not reuse enough ones that. to the. we. knowledge from past decisions.. or other groups have. more groups make decisions about that. referred. I. to.. the. made. before.. We often make decisions For example,. computing environment since. were shared among these groups so. knowledge accumulated by one group could have been useful. The reusable knowledge includes more than. have helped three. the original experience. There were enough requirements and alternatives. interface to email) that. I. similar. that. minimize. (e.g.. many. cost,. pieces of the. to the others.. factual information; the. knowledge from past. decisions might reveal a critical requirement that current decision makers did not think. about, an option they were unaware of, an assumption they mistakenly held, or an. argument they did not consider.. Also valuable. is. the. knowledge about how different. requirements relate to one another; for example, the knowledge that minimizing cost can typically be factored into minimizing purchase cost, minimizing maintenance cost, and. minimizing development considered. cost.. If. in similar past decisions. nothing else, the knowledge of the requirements. can serve as a useful checklist.. knowledge from past decisions, however, the. is. usually ignored.. As. Much. a result,. of the useful. we. often repeat. same effons and mistakes.. Moreover,. we need. to. understand the rationales for a past decision for reasons other than. reusing them in similar decisions. Another problem in this category audit. trail. so that. we can justify. or review our decisions.. where. the rationales for the original design. of the. artifact.. In all. of these cases,. if. is. that. of keeping an. Yet another problem. is in. design,. can be valuable for troubleshooting or redesign. we had some way of. relevant parts of the deliberations underlying past decisions,. selectively accessing the. we would be. able to learn.

(21) much from them,. not to mention saving the effort of collecting those pieces of knowledge. in the first place.. Summary. 1.2. In the last section,. of. Approaches. described and categorized the problems that motivated our research.. I. propose a Decision Rational Management System as a solution to these problems. Decision Rationale Management System (. 1. (RMS). I. A. has three components:. a language for describing the elements of decision rationale,. (2) a. method. (3) a set. for using the language to capture the rationales,. of services that use the capmred rationales to provide decision support.. For example, a simple example of decision rationale management system. is. to. (1) use English,. (2). wnte down everything. (3) find. said. by anybody. in the. decision. making process, and. whatever we need from the record by brute force. (e.g.. by flipping through. pages of the notebooks used).. This solution later finding. is in. what. most cases unsatisfactory. The cost of writing down everything and. is. needed. is. so large that the benefit of actually finding something. is. not. likely to override the cost.. So we need. to design a decision rationale. decision rationale management,. whose. management system,. benefits exceed. challenge in the design of a decision rationale (1). design a language that. is. structured. design rationale and their relations.. its. costs.. management system enough. i.e.. a tool for supporting In other words, the. is to:. to capture the. important elements of.

(22) (2). develop a method which helps reduce the cost of using the language. to. capture. rationales, and. which reward the user for recording rationales.. (3) define services. A rational management system the. problems discussed. in the. (RMS)'. previous section.. developing a system called SIBYL. characterize the. capture rationales in. 4).. DRL,. is. based on a model. SIBYL. provides an environment. et al. 1988], a tool for building. in. which. This environment. is. SIBYL. rationales (Chapter 6).. from past decisions,. In the rest of this section,. built. make. also provides a set of computational. services, such as keeping track of dependencies,. retrieving relevant rationales. to. computer supponed cooperative. features such as template editors and various display formats, to. the rationale capture easier (Chapter 5).. that. rationales (Chapter 3), and provides. DRL as a by-product of making decisions.. on top of Object Lens [Lai its. them (Chapter. can solve or alleviate. This thesis substantiate this claim by. language,. Its. imponant elements of decision. constructs for representing. work, and uses. that satisfies these constraints. I. comparing multiple versions, and that. reward the user for recording. provide an overview of. how SIBYL. addresses the problem categories discussed above.. Managing Rationales within. As mentioned. in the. a Session. previous section, there are. decision analysis, help. manage. many. decision suppon tools, such as. rationales, but their use requires the precision or the. formalization often not available or too expensive to produce in. '. In the rest of the thesis,. I. will often abbreviate Decision Rationale. Management System.. 10. many. decisions. In order. Management System. as Rationale.

(23) manage. to. rationales in these situations,. representation (Chapter. 3).. That. is,. the constructs of. rationales, such as alternatives, goals,. formal types with their. own. SIBYL implements DRL. as a semi-formal. DRL for representing. the elements of. and arguments about them, are implemented as. attnbutes, but allows the values of these attributes to be a. mixture of formal and informal descriptions.. This way, the domain knowledge can be. entirely described informally (as attribute values of the formal constructs), but. still. help. manage. For example, English, but. of the of. all. on the way. rationales based. that the formal constructs. the information, say about an interacuon. SIBYL. can provide. DRL constructs. because. its. service as long as. knows how an. it. it. knows. SIBYL. can. have been used.. manager, may be given that. it is. in. an alternative, one. alternative should relate to other constructs. DRL.. Using the rationales captured. managing. among. rationales within a session.. SIBYL. arguments. that. depend on. DRL) it.. is. The. no longer. provide the following services for. allows the user to maintain the consistency. For example, when an assumption. the elements of rationales (Section 6.2).. (represented as a claim in. for. DRL, SIBYL can. in. true, the user. can ask. SIBYL. user can also create and compare multiple decision states,. example, a given decision under different assumptions (Section. structure of. DRL, SIBYL can. for example, all the. evaluations.. suppon. tools. knowledge.. are not as powerful as those provided by. the other hand,. domain knowledge, and services.. alternative, or the criteria. and often require interaction with users. On. 6.3).. Using the. also keep track of the rationales by collecting and displaying,. arguments evaluating a given. These services. to invalidate the. SIBYL. in the. some. used for the. other decision. absence of formalized domain. allows the user to create formal objects modeling. write rules that exploit the. knowledge. to obtain. more powerful.

(24) Managing Rationales across Sessions. Because the rationales are captured rationales across sessions. permanently available. becomes. in. during a session with SIBYL, managing. easier for several reasons.. in the electronic. because they are represented. DRL. form accessible. in a structured. way,. SIBYL. First, the rationales are. to the computer.. Furthermore,. can easily find decisions. that are. unresolved from the previous sessions, questions that need yet to be answered, or the. arguments that need yet sessions by. making. provides a forum. in. to. the. be evaluated.. SIBYL. in fact helps. boundary between the sessions. which. to. examine and update the. managing. decision without having to be physically together (Chapter. decisions asynchronously also solves. some of. the. Because. less rigid.. rationales,. 5).. problems. its. rationales across. users can. SIBYL make. This ability to. in face-to-face. a. make. meetings. mentioned above: such as delay caused by meeting arrangements.. Reusing Rationales across Decisions. In addition to the benefits of providing a. past decision rationales,. SIBYL. permanent, electronic, and structured record of the. provides services that specifically help the user to retrieve. relevant pans of the rationales from past decisions. rationales e.g.. once they are represented. in. DRL.. If the. There are two ways of retrieving the user. is. looking for something specific,. arguments about a given alternative or answers to a question, then the user can specify. a partial structure as a query.. from past decisions. that. A. matches. rule. system takes. this query,. and retrieves any structure. this partial specification (Section 6.. 1. 1).. Users, however,. are often interested in relevant rationales without necessarily looking for anything specific.. To. support such cases,. I. have also developed an interactive algorithm for retrieving relevant. 12.

(25) rationales for a given decision (Section 6.1.2).. manager,. is. This algorithm, labelled the precedent. based on the inmition that two decisions are similar to the extent. that they share. similar goals, and uses the goals shared between a given decision and past decisions to. judge potendaJly relevant rationales.. Sharing Rationales across Decisions. Sharing rationales becomes easier again because their relations explicit.. altemaDve or answers to the. DRL makes the elements of rationales and interested in all the arguments about an. For example,. if. the user. quesdon.. is. easy to collect the relevant elements and send them. to a. it. is. group. Using the features of the underlying Object Lens, mentioned above,. can send and receive a collecdon of structured objects through email or. them appropriately. to the objects that already exist in the current. in a file,. SIBYL. and. link. environment (Chapter. 5).. This way, a group can request and access pans of rationales of the other groups and use this. knowledge. communicadon. as a basis for mutual understanding, for collaboration, or for further. across groups.. In the next chapter,. I. describe a scenario that illustrates the overall idea from the user's. point of view. This scenario illustrates the benefits that will motivate the user to use the. language and the kinds of services that an. RMS. 13. needs to provide..

(26)

(27) Chapter 2. A. Scenario. This chapter. illustrates the. behaviors of an idealized decision rationale management system. which the implemented system has. tried to. approximate. This scenario represents a. step taken, logically as well as historically, in the present research.. Logically,. it. first. specifies. the goals of the research in concrete forms. Historically, the rest of the research discussed in later. chapters. services. -. --. follows the articulation of this scenario.. to present the goals. interface. to be. design and implementation of the language, the interface, and the. It. As. such, the scenario provides a. way. and main ideas clearly without getting into the details of the actual. also serves as a yardstick against which the success of the current research. measured.. 14. is.

(28) The system presented below ways.. First,it. system, makes. different. it. natural language, though. implemented system. beyond. point the details of the. at this. buttons that the user needs to activate.. described in Chapters 5 and. the. the. in the. following. the capability of the. implemented. possible to describe the user request and the system response without. having to describe. at this. from. uses natural language interface in order to get the main ideas across more. The use of. clearly.. is. 6.. The. SIBYL. interface, such as the. actual interface of the. implemented system. are slightly different. is. Also, in order to present the main goals and ideas concisely. point in the thesis, a step in the scenario sometimes represents a. The scenario uses. implemented system.. menus and. a. few features, such as multiple shading of the nodes. when. steps in. few terms, such as issue and subissue,. from the ones actually implemented but more. scenario, though not implemented,. number of. in a. that. A. intuitively obvious.. graph, are presented in the. they help getting across main ideas better.. The. scenario therefore should be regarded as a concrete illustration of the goals and the main ideas that have guided the present research not as a description of the current system.. Imagine a group designing a window manager,. manages windows. i.e.. the software. for different applications. Typically a. window manager would provide. functions that an application can call upon, for example, to resize,. and. scroll a. window or. There are different ways. ^. in. structure the. which provides and. do something. components of a window. which these functions can be provided. like create,. in a. move,. cenain way.^. to the applications.. At. Window manager. we. design was chosen as an example domain for several reasons. First, it is a domain that beginnmg to understand. The domain knowledge has not been formalized, but parts of it could knowledge about modularizauon. Thus, it provides an opportunity to explore the feasibility and. are just. be, e.g.. the usefulness of a rationale. domain faces. management system. in realistic situations.. Also,. many. issues that this. are potentially reusable for other domains. For example, the issue of whether to have a suict. between modules is quite general and anse m many other domains,such as the design of operaung systems or programming language. Thus, it provides a chance to explore cross-domain reusability of the rauonales as well, which is not discussed in this thesis, but is a topic of future research. In constructing interface. the scenario, [Lane 1990]. and fHopgood. et al. 19851. have been very useful.. 15.

(29) one extreme, applications can have access any of the window manager functions. other extreme, applications can have no access to any of the directly but. interface. of. have. to. go through a layer of abstraction. management. window manager. direct access to the. (i.e.. window manager. functions. an interaction manager or the user. system), which typically consists of higher level routines. functions.. At the. composed. There are also intermediate solutions, such as allowing. window manager funcdons. but also providing a toolbox (or library) of. higher level routines that applications are encouraged to use.. In the scenario,. two. alternatives are currendy being considered: the interaction. the toolbox mterfaces.. scenario stans. at a. These alternatives are described further. point where the group has used. accumulated some arguments concerning them. group,. is. about to use. to contribute additional. SIBYL. to catch. SIBYL. to raise a. Imagine. up on the current. in the. scenario. itself.. number of. that John, a. state. manager and. issues and. member. knowledge.. User's requests and the system responses are in the courier font. rê actions taken by the user or the system are in italic courier. the figure captions are in Times.. Getting Background Knowledge. User:. Show me the current issues,. 16. of this. of the decision making and. Notation:. The comments and. The.

(30) Target System: displays graphically the issues and their relations.. How. {Fig. 2.1). best to design. WING,. How ponable should be our. our influences. window manager?. applications?. subissue subissue. Should windows. What kind of interface for. be supponed. applications?. across network?. subissue subissue. Should. WING. suppon. direct manipulation. What kind of. style?. knowledge should be visible to the. applications?. Figure 2.1 A browser displaying the issues currently being considered. The ones heavier boxes represent unresolved issues.. 17. in.

(31) User:. double mouse-clicks on the issue,. "What kind of application. interface?" to get more information about it.. Target System: displays all the information about the chosen issue in the form of attribute/value pairs.. (Fig.. 22). There are two. alternatives being considered: Interaction Manager and Toolbox approach,. Issue:. as shown in the Alternatives attribute.. What kind. of application inierface?.

(32) ALTERNA. 1. VE: Interacuon Manager.

(33) as a result. of these changes.. From. these evaluations, the user can trace the changes. responsible for these updates, as illustrated in the scenario below.. however, the system can also replay the changes. change can be displayed, for example,. User:. at the user's. in the. Optionally,. chronological order; each. mouse-click.. Give me the updates.. Target System:. The following describes the objects added or. changed since the last time you used the system that might be relevant to the current decision.. You might want to examine them. before you proceed.. The decision problem,. [Should the windows be supported across. network?], which is. subdecision of the current one, has been. a. resolved in favor of the alternative,. [YES,. but only local. network]. The decision problem, to applications?),. [What kind of knowledge should be visible. might be worth reexamining because an. assumption on which it is based has changed.. The system can be. told to. monitor a particular change and, when the change occurs, do. certain things, like propagating. its. consequences and/or notifying the user. The price of the X terminal has just been reduced.. changes the evaluations in the decision problem, windows be supported across network?]. 20. .. That. [Should.

(34) User:. Show me the current status of the issue,. "What kind of. application interface?". Target System:. displays. a. matrix that shows current evaluations of. the alternatives with respect to different criteria.. The Current. Status of Decision Problem:. (Figure 2. 4). What kind of application. interface?.

(35) User:. Show me the reasons for the evaluation of Interaction Manager. with respect to Reduce development cost.. Target System: displays for the evaluation.. a. browser displaying arguments responsible. (Figure 2 J). Arguments evaluating "Interaction Manager" with respect to "Reduce Development Cost".

(36) the current implementation does not. suppon such. a feature.. The nodes shaded. are the. ones that have remained unchanged.. The user could. also find out the reasons underlying other evaluations by. argument browser associated with each of the evaluations. Also,. if. examining the. the user does not. understand some of the arguments, the user can get more information by expanding the. nodes. For example, suppose. in question.. qualification. was. in the. that. John did not understand what the. above argument browser.. User: double clicks on the qualifying claim to expand and examine it in more detail.. Target System:. CLAIM:. This. dvpmt. costs.. displays the qualifying claim in full.. is. one time. cost.. Once. built,. it. will. (Fig.. reduce applicadon. 2.6).

(37) Making Changes. User:. enters additional arguments, and replies to question raised. after examining the current state of the decision.. Target System: helps the user to update the argument browser through context sensitive menus and displays the updated browser (Fig. 2.7). Arguments evaluating. "Interaction. Manager". with respect to "Reduce Development Cost". Interaction. Manager. ACHIEVES Reduce Development Cost. denies. This. is. one time. Once. built,. it. cost.. will. reduce appiicaiion. dvpmL. We built one before, though for a different. answers. platform. 1 Has not been. finalized,. but around $. Figure 2.7. costs. 40K. The argument browser updated with. 24. additional arguments..

(38) The. user can also relate. one claim attribute to. to. two claims with stronger-than. (or equal-to) relation if. he judges. be stronger than the other claim. The user would also have to specify the. on which they are being compared. compare any two objects of. sometimes figure out from. this. the. --. e.g. plausibility.. same type using these. In fact, the user is able. relations.. The system can. comparative information which alternatives dominate. others.. The consequences of. For. additional claims or questions entered are propagated.. example, the decision matrix shown earlier would. now show. appropriately depending on the type of changes (e.g.. marked. the relevant cells. marked. for reevaluadon).. Reusing Past Decision Rationales. User:. Find me useful information from past decisions. Target System:. The following decision matrix summarizes the. updated evaluations, given the new knowledge transferred from past decisions.. The evaluations are tentative,. subject to your. modifications and confirmations."' (Fig. 2.8). is shown as responding immediately here, the implemented system requires interacuon with the user lo produce this response, as discussed in Section 7.1.2.. ^. Although the system. 25. much.

(39) The Current Status of Decision Problem:.

(40) Interaction. Manager achieves the goal,. ^. [Implement direct. IMcanbe. denies. extended. manipulation]. qualifies. Has only limoed. denies. knowledge. comments. to. support applicaaon-de fuieddata types. *"/\. of Its. style is too rigid. because he used an. to support direct. semantics. denies. \ Some. manipulation (Myers 87). interacuon. supports. application. interaction. Ask Sam. It is still. manager. direct. hard even if applicauon. T. manipulabon this way, eg-. defined types are. denies. IM's. suppon. MacApp. allowed. raises. In our case, it does not matter because we have only simple objects to manipulate. answers. through procedures. Figure 2.9 An argument browser displaying the aiiUments incorporated from past decisions, potentially relevant for the evaluation of the alternative. Interaction Manager, with respect to the goal, Implement direct manipulation.. The. user can examine the. modifying them middle, "Ask. manager. is. Sam. it. objects, deletes those not useful,. necessary.. and keeps. the rest after. For example, John deletes the argument on the. because he used. a context-sensitive. decision context. case,. if. new. ..". because the fact that. knowledge. that. Sam. used an interaction. does not carry over. Also, John modifies the argument on the bottom. to the current. left side, "In. does not matter because we have only simple objects to manipulate," to. 27. left. "It. our. may.

(41) maner. not. if. we. decide not to allow customizable icons." In other words, the original. argument did not exactly carry over, but suggested something close. that. is. relevant.. Managing Multiple Viewpoints. John wonders how evaluations would change. if. he gives up the goal of ponability.. Suppose we do not consider [Portability]. User:. .. How does that. change the evaluation?. shows. Target System:. The user can examine. a. the reasons for the. Now. objects have been marked.. back to. User:. It. later or. compare. Create. a. Target System:. changed decision matrix. it. (not. shown). changed evaluation. As before, the changed. he wants to save the current state so that he can. to other states that. come. have also been saved.. called "Without portability".. viewpoint,. The new viewpoint will inherit from the initial. viewpoint, and the changes that have been introduced since the initial viewpoint are:. •. •. [Portability]. The following objects have been added: (GOAL),. •. is no longer a goal.. [Common User Interface]. [Extensible Interaction Manager]. (ALTERNATIVE),. The following objects have been inactivated: Manager]. (ALTERNATIVE). ,. .... 28. .... [Interaction.

(42) The following objects have been changed: the importance. •. attribute of [Portability]. Is it o. User:. .. )c.. (GOAL),. .... to save this new viewpoint?. o.k.. Target System:. The viewpoint,. "Without portability," has been. saved.. Wrapping Up. What do we need to do to resolve the current decision?. User:. System:. Target. •. We need to resolve the decision,. [Should the windows be. supported over the net?]. •. We have to answer the Question,. [What are the examples of the. interaction manager based systems and their experiences?]. John continues. to. examine. manner described above.. User:. Let. rr.r. the decision problem,. <now when the decision,. System:. o. .. k. alternatives, arguments,. in the. Finally, he asks the system to monitor a decision for him.. supported over the net?],. Target. its. .. is made.. [Should the windows be.

(43) Chapter. 3. The Structure. of. Decision Rationales. The goal of imponant. this. chapter. is to. characterize the elements of decision rationales that are. to represent explicitly.. task for which the representation. have been. Of course, what is. used.. illustrated in the scenario.. form of questions.. Then. I. tasks for a rationale. In the first section,. make up. more complex because. from one model. important to represent depends on the. I. management system. summarize these tasks. to the next.. decision rationales.. the objects. and. For each model,. I. The consecutive models. (Chapter. DRL. (Chapter. 4), to. are. relations are successively differentiated. discuss what extra tasks the additional. refinement allows us to do. The framework provided by these models the thesis, to present. in the. develop a sequence of models, each of which makes explicit the. objects and the relations that increasingly. The. is. is. used. in the rest. of. discuss the kinds of rationales that can be reused. 6), as well as to discuss related representations. 30. (Chapter. 7)..

(44) 3.1. What do we want. One way of. to. characterizing a task. accomplish the. do with decision rationales?. is. to. list. the questions that. Given the goal of producing the behaviors. task.. we need. to. answer. to. illustrated in the scenario,. our representation should be able to answer the following questions, abstracted from the scenario.. What. is. What. did. What. are the alternatives being considered?. the status of the current decision?. we. discuss last. week and what do we need. to. do today?. Wliat are their pros and cons?. Why. do we even consider. discussed. What. How does What. Why. if. two most favorable. this. we do. is this. and how. new. not consider this goal?. anyway?. are the decisions that. What. are the unresolved decisions?. depend on. decision?. this. What. are. we. What's the consequences of doing away with. How did other people deal. list. learn. from. with. this. currently doing about them?. this. assumption?. problem?. the past decisions?. of the questions. illustrate the differences in. are,. we. alternatives so far?. What. The above. related to the one that. fact affect the current evaluations?. goal important. What can we. is it. week?. last. are the. this alternative,. is. by no means complete.. They have been chosen. what the following models do or do not allow us. however, representative of the questions that arose. 31. in. to do.. to. They. our experience of capturing and.

(45) managing decision. As. rationales.. of a model can be checked.. If. such, they serve as test cases against. answered by a particular model of decision a task of the. system based on. some questions which cannot be. turns out that there are. it. model or. that. rationale, then. the system. which the adequacy. answering those questions. would have. to. extend. its. not. is. model.. Models of Decision Rationale. 3.2. We now. develop a series of progressively richer models of decision rationale. As shown. in Fig. 3.1,. rationales.. each model makes explicit the objects and the relations that make up decision. The successive models. are increasingly. more complex because. the objects in the. next model in the sequence differentiate the objects in the previous one.. Decision rationale. made. the. way. it. in the. was. So in our. with a body of reasons as. be. made. earlier.. relations. first. shown. An example. We. is. is. an explanation of. model of decision. in Fig. 3.1a.. explicit to different degrees.. undifferentiated.. used. most general sense. The. rationale, an anifact. is. was. associated. At one extreme, they can be completely. the natural language description of a historical record. we. can also imagine a representation, however, where logical suppon. among reasons. to refer to. the decision. internal structure of these reasons can. is. made more. explicit. by providing constructs. Implies, Supports, Denies, Qualifies, and Presupposes.. Space. why. We. will use the. like. term. Logically. Argument. what we have called a body of reasons because the reasons are captured. either as a historically recorded or logically structured record of the various arguments. relevant for a decision.. 32.

(46) (a). MODEL. 1. :. A. decision outcome. is. Z":?*^^ 'O-.l,. associaicd with a body of. all. the arguments e:^plaimng the outcome.. /ITBI,, 'N*T,vj. «lt«T,. •a iH'"^! -'*n-^''*'»Qk„ "•0«r. 'DOr,. oach). (b). MODEL 2:. Aitemaaves and then. relations are. made. explicit. and. the. arguments about individual alternatives. can be differentiated.. /. MODEL 3: Evaluation measures used and then- relations are made explicit and the arguments about them can be diffeientiaied. (c). Figure 3.1 (a)-(c). Models of decision rationale (continued on the next page). 33.

(47) (..^^•^. CHlTt. "'"î*,. '"ô,.**^. (d). MODEL 4: Cntena used for evaJuaaons and their relauons are made explicit and the arguments about them can be m the argument space.. further differenoaied. •«u»i. "^""^ '*«*!,. •"Ml (e). MODEL 5:. discussmg the. Individual issues are issue.. made. explicit each of. which contains the altemauves. evaluations, and cntena used the meta-arguments about the issues and their relauons.. A pan of the argument space includes. Figure 3. 1 (d)-(e). Models of decision rationale (continued from the previous page). 34. :.

(48) There. on. much we can do. is. this. with our. model can help us answer. do we need. to. do today?" Such. How did other people deal Our. first. first. model of decision. a representation. much. qualify this statement immediately. Saying that. cannot answer these questions.. on the model has language free. model. more. itself. all. text, these. it. learn. is. we. shall. we. not saying that. if. So. it. is. just in the. the real issue. is. form of natural. how much. the. directiy.. We. by providing computational services. will see. how more. differentiated. that help us. models allow us. easily, although they increase the cost in. to. some other ways. & Yakemovic in print].. (Fig. 3.1b) differs. from the. first. by making explicit multiple alternatives. Decision making involves formulating several alternatives for a. their relations.. problem, comparing them, and merging them as needed. In our is. made. explicit at a given time,. and. all. first. ones that have been rejected. Once the alternatives become. explicit,. attributes (e.g. current status such as "rejected" or "waiting for. the relations. among. that specialize another alternative), or. The. explicit, including the. we can. talk. about their. more information"), make. the alternatives explicit (e.g. specialization, historical precedence),. define computational operations on them. considering.. model, only a single. other alternatives are present only implicitiy. argument space. Our second model makes the alternatives. many of. the past decisions?. helps us answer the questions by making the structure of the argument space. Our second model. in the. from. does not help much. questions can be answered.. answer these questions more. solution. week and what. user works hard enough and the representation based. explicit for us to reason about or. [Conklin. last. with the other questions, though. necessary information captured, even. answer the questions. and. If the. representation based. can also help us answer the questions:. What can we. with this problem?. A. "What did we discuss. the questions,. model, however, does not help us. rationale.. (e.g.. compare. alternatives, display the alternatives. even argue about whether an alternative. is. wonh. alternatives other than the one finally chosen are interesting because. the issues discussed. and the knowledge used. 35. in evaluating. them. are. imponant. in.

(49) other contexts or for re-evaluation of a discussion. Space. use the term Alternative. when. to refer to this set. We. situational constraints change.. of multiple alternatives and their. relations.. The. relations. among. alternatives can be historical or logical. Historical relations. only the linear sequence that relations such as layers. may. we. usually describe as versions, but also. and contexts [Bobrow. &. [McKinlay. et al.. 1990].. To. Or. is. the extent that. we. we want. a. say that the. within the scope of the representation.. By now, we have an 3.1b.. more complex. alternatives can be. representation to represent different alternatives and their relations, alternative space. be not. Goldstein 1980]. The logical relations. include Specializes, Generalizes, Elaborates, or Simplifies.. related through a design space. may. alternative space connected to the. argument space, as shown. in Fig.. For each of the alternatives, there are arguments describing the reasons for. its. current evaluation, just as in our. first. model. there are arguments describing the evaluation. status of that single alternative,. i.e.,. that. was chosen. Some of. it. the arguments can be. shared; for example, an argument can support an alternative while denying another; so better to think of the. arguments about the different alternatives forming a single large. argument space, as shown. in Fig. 3.1b.. Once. is. the alternative space. helps us answer. it is. represented,. some new questions. in. we can imagine how we can make our. list.. To answer "What. being considered?" and "What are their pros and cons?,". we can. a system that. are the alternatives. associate an argument. space with each of the alternatives through relations such as Supports or Objects To. representation of the alternative space structural relations (e.g. Is. A Part. makes. Of),. explicit historical relations (e.g. Replaces) or. among. 36. If the. the alternatives, then. it. can also help us.

(50) answer questions such as to the. one. that. we. "Why do we even. discussed. last. consider this alternative, and. explicit inultiple alternatives,. what the argument space. is. about. In our. first. we need. to articulate. model when we had a. more. Similarly, the arguments for the other alternatives are about. why. why. carefully. single artifact,. chosen one, the argument space contained reasons for the choice of. to generalize,. related. is it. week?". However, once we make. the. how. namely. that artifact.. they were not chosen, or,. they have their particular evaluation status, e.g.. "Still in. Consideration",. "Waiting for More Information", "Rejected". 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,. we. new. introduce a. evaluation statuses of. alternatives are. all. made. Often, the implicit ordinal relation. the alternatives.. is. sufficient;. However,. these values, for example, a. measures. to. if. among. we. we want. program. leave. and. explicit. not and need not specify any elaborate relation. "Poor," "Very Poor"). Space. space. Evaluation. among. to the. to define. We. need. human. values to produce a higher level. mean so. that their interpretation. the. human. 37. use.. to be. that. manages. and merges evaluation. very careful about what. measurement, some calculus for. user. summary measure, we need does not become. we. user to assign these values to. combining them, and a model specifying what they mean. Even if. we do. "Very Good," "Good,". that automatically propagates. actions are left to the user, for example,. Usually,. any computational service. to specify the units of. where the. the evaluation measures. produce a higher level summary, then we need. these values mean.. inter-related.. these values (e.g. it. (Fig. 3.1c),. arbitrary.. is. in the case. expected. to set. to. where these. combine these. down what. these values.

(51) Making. the evaluation space explicit allows us to differentiate. argument space: and. arguments about. (1). why an. alternative has. arguments about the alternatives themselves,. (2). even consider an object as an alternative or whether. With a representation of. another alternative.. e.g.,. two components of. current evaluation status,. its. why we. should or should not. this alternative is really a special. the evaluation space,. new. We can also explain. fact affect the current evaluations?". case of. we can now answer. questions such as: "What are the two most favorable alternatives so far?" and this. the. "How. does. an evaluation measure. pointing to the arguments in the argument space behind the decision in question, and elaborating on. how. arguments or how. Our models However,. the particular evaluation. so far do not. to represent explicitly.. evaluation.. explicit the criteria used in producing an evaluation.. and. For example,. is. it. window manager and. By making. their relations are usually quite important. important to. application". Manager' because o/the goal of. this criterion. important,. make. the criteria used for the evaluation. "Interaction. derived or computed from these. is. related to other measures.. it is. separation between. measure. this criterion explicit,. we can group If. appropriate operations on. which all. knowing. that. We. use the term Criteria. Fig. 3. Id shows,. once. we have. used as a criterion for. arguments. all. to. Space. change. Knowing how. its. importance. to refer to these criteria. the criteria space explicit,. we. is. 38. a. way. when and. becomes. less. a criterion. is. of achieving. related criteria. their relations.. As. can further differentiate the. argument space by grouping those arguments which are about the relations.. that appeal to. the arguments that presuppose this. "Implement direct manipulation". "Naive user support"), also allows us change.. the. is. the criterion changes or. goal (for example, making these arguments less important). related to others (e.g.. argument "Clean. that the. a pro-argument for the alternative. portability,. and weigh them against one another.. we can perform. is. know. criteria. and. their.

(52) Hence,. is. it. imponant. language whose scope includes the. that a. different attributes of the criteria. criteria. space represent the. and the relationship among them. For example,. should. it. allow us to represent the importance of these criteria and the synergistic or tradeoff relations. among them.. A set of criteria can. facilitates the satisfaction. in. of the. be sub-criteria of another. latter.. in the. sense that satisfying them. These sub-criteria can be related among themselves. various ways: they can be mutually exclusive in the sense that satisfying one. makes. it. impossible to satisfy others; or 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 their parent criterion in various ways as well: they can be exhaustive in the sense that satisfying. all. of them. is. equivalent to satisfying the parent goal; or they. may. not. cover the parent goal completely.. With. the criteria space represented,. us answer questions such as: what. we can now if. we do. important anyway? The answer might be "If evaluation of the alternative. X. X. changes. to. see. how. the system. not consider this goal?. we. might be able or. why. "High" because. all. is. imponant because. representation of the criteria. is. if. would have. we want. to provide. far,. we have. artifact,. criteria,. and. seems a necessary. these questions.. At. least,. we. the information necessary to define an operation that will give or suggest the. We will. answers to these questions.. So. some support answering. a. not sufficient for answering. these questions; however, the explicit representation of the criteria space. condition. it is. These answers can be. derived only from a representation that makes the relation between evaluations,. Of course,. goal. these claims that argue against. subgoal of another important goal. Have a wide distribution.". explicit.. is this. give up this goal (say portability), then the. were based on the imponance of ponability." or "Ponability. arguments. to help. identified. give examples of such operations later in Chapter 6.. and discussed the structure of a single decision underlying an. namelv which of. the alternatives should. 39. we choose?. However, with. the.

(53) representation of such local structures alone, namely. and. criteria space,. we. still. the unresolved issues?". such questions,. cannot ask some of the questions. and "What are the issues. we need. more global. a. can be a sub-decision of another. if. can be a specialization of another. imponant. Space as. its. to refer to them.-^. A. many. if. issues and. we. we can. still. unit in this issue space. answer questions. shown. some questions. that. we have. we. a. more. are. A. is,. A decision. detailed case of the second.. we. use the term Issue. therefore, a single decision that has. in Fig. 3.1e.. Once we have an. "What. are the issues that. not yet covered such as: learn. from. issue as. Taken" with. depend. How. enough information. "What. among. the. this issue?". did other people. the past decisions?". far identified, the spaces of. evaluations, criteria, and issue, can contain 6,. is. the first decision.. are the unresolved issues?" as well as. "What can we. however, that the five spaces so. Chapter. making. Representing the dependency relation like. "What. To answer. this issue?". associate attributes such as "Status" and "Actions. deal with this problem?" and. In. as:. individual issues are related.. these decisions are related, and. currently doing about them?". There are. such. other decisions. For example, a decision. the first decision. answer questions such as "What. issues will allow us to. in the list. depend on. how. the laner requires. internal structure the other spaces, as. an explicit element,. are. how. to capture. that. picture of. decision often requires and/or influences. It is. argument space, alternative space,. its. We. argue,. arguments, alternatives,. to. answer these questions.. discuss computational operations that help us answer these additional. questions by exploiting the structure of the five spaces.. '^. The temi Issue Space used to refer. U) the set. is used instead of Decision Space because of ail objects relevant lo making a decision.. 40. Decision Space. is. someumes.

(54)

(55) Chapter 4. DRL:. A. Decision Representation. Language. This chapter presents a language, rationales.. developed. DRL in the. DRL, developed. for representing. has been designed based on the last model in the sequence of the models. previous chapter.. As. such,. it. provides constructs for representing the five. spaces of argument, alternative, evaluation,. criteria,. The philosophy underlying. DRL. spaces,. DRL. imponant. to. the design of. and. issue.. has been a minimalist.. For each of the. staned out with the fundamental object type and the relation types that are. answer the kinds of questions enumerated. example, for the. and the. and managing decision. criteria space,. relation, IS. DRL provides the object type, GOAL,. A SUBGOAL OF,. portability important?. in the previous chapter.. which. is. 41. to represent a criteria. necessary to answer the question,. In the course of using. DRL. For. Why. is. to represent decision rationales..

(56) however, additional questions arose.. answered by the existing constructs. shows. the constructs that. all. And. is. it. only when these questions could not be. that additional constructs. DRL currendy provides.. were added.. Although not large. Figure. 4.. number, these. in. constructs have proved to be adequate in representing the cases and providing the services that. I. we have. first. far.. give an overall description of. map. constructs their. explored so. to the five spaces.. DRL. More. and then discuss,. detailed descriptions of the. semantics are provided in Appendix. descnbes. DRL. language. may seem complex,. in the next section,. I.. I. would. like to. DRL. emphasize. much of these. these. constructs and. that this chapter. as a language, not the user interface for using the language.. but. how. At times, the. complexities can be hidden from the user. through user interface such as context sensitive menus that display appropnate actions in appropriate contexts.. 4.1. The. actual user interface. is. described in the next chapter.. Overview. This section provides a high-level overview of constructs. DRL. DRL, and. the next section discusses the. provides for representing each of the spaces discussed in the. Figure 4.1 shows the object types that form the vocabulary of. Relation and. its. DRL.. subtypes can be used to link two other objects.. objects can be used to link an Alternative object to a. Goal. object.. last chapter.. Objects of type. DRL. For instance. Achieve. The. legal types that. can. be linked are shown inside the parentheses following the names of the relations. Figure 4.2. shows. the. schema. rationales at a given. for a decision graph,. which. is. moment. This schema shows how. used the. to graphically illustrate the. DRL constructs. are related to. one another. Figure 4.3 shows a decision graph for an example decision problem.. 42.

(57) Altemaove. Achieves (alternative, goal). ,CkMl ^-^—'^^^—. -. Decision Problem. Is. a. Good Altemaove. for (altanaave, decision. problan). Supports (claim, claim) Denies (claim, claim). Presupposes idaint, claim) Is. A. Subgoal Of (goal goal). — •. Claim. Is. Is. a. Subdeosion of (deosion problem, decision problem). ReUtedTo. DRL Answers. Obiect. (claim, question). Is. An Answering ProoedureFor. Is. A. Result. (procedure, question). Of (daur. procedure). Tradeoffs (ob|ect, object, attribute). 'Question. Is. A. Kind Of. (ob|ect. ob|ect). <. Group Viewpoint. Raises (object quesbon). Omments. (claim. ob)ect). Procedure Status. Figure. 4.. 1. The. ^— Deaded. DRL. Vocabulary. ([Go^ IS. a subgoal of I I. IS. a good ailemauve for. AJtemaave. achieves. denies. IS. a. good altemauve. for. /. IS an answering procedure. Claim. for. denies. /. presupposes. Gaim IS. a result of. Claim. achieves. achieves. Altemanve. Figure 4.2 The schema for a decision graph, showing the relationship major constructs of DRL. 43. among. the.

(58) Figure 4.3. An Example. Decision Graph in. DRL. 44.

(59) A Decision Problem represents place the. window commands. An. A. [EM] or [Toolbox].. is. is. its. that requires a decision; for. subgoals.. example, where. to. Alternative represents an option being considered: e.g.. its. In panicular, a decision. subgoals.. X for Y,". of the form, "Choose the oprimal. i.e.. comparing the. a desirable state or property used for. elaborated in terms of. a goal of special kind,. elaborated by. is. problem. Goal represents. A Goal. alternatives.. problem. the. where. Y. For example, choosing the best alternative for the decision. problem, [Which application interface for WING?], means choosing an alternative that satisfies as. UI. styles].. much. as possible. Every relation. its. in. subgoals: [Reduce dvpt cost],. DRL. example, the rightmost Achieves link. is. [Is portable],. a subclass of Claim, as. in Fig. 4.3 represents the. shown. Claim. [Suppon many. in Fig. 4.1.. For. that the Alternative. [Toolbox] achieves the Goal [Implement direct manipulation].. We. evaluate an Alternative with respect to a. between the Alternative and the Goal,. We. i.e.,. Goal by arguing about. the claim that the Alternative achieves the Goal.. argue about a Claim by producing other Claims that Support or. qualifying the Claim by pointing out the Claims that. of an alternative. is. the Achieves relation. it. Deny. is. a. Claim or by. Presupposes. The overall evaluation. represented by the Evaluation attribute value of the Is. Alternative for link between the Alternative and the Decision Problem, the alternative. the. good. i.e.,. A Good. the claim that. alternative resolution for the issue. This evaluation is a function of. the evaluations of the Achieves claims that link the Alternative to the different Goals.. The constructs of the five spaces. space. in. DRL. DRL can. and those. that are generic.. The constructs discussed so. far are specific to a. also provides constructs for representing objects and relations that are useful. any of the spaces. Group. "Member. be divided into two major categories: those specific for each of. is. used to represent a collection of objects and has the attribute,. Relations," which tells us. how. the objects are related.. of objects rather than a single object. For example, a Goal. 45. A relation can. may. take a. be related to a. Group. Group of.

(60) other Goals through a fs. A Subgoal of link. The. usual generalization (or specialization) relation. next section and in Appendix. 4.2.. each of the five spaces. The other constructs. Of, represents the. are discussed in the. of the Decision Rationale. DRL constructs. in the. A Kind. I.. DRL's Representation. This section discusses the. construct. Is. model. that. in. more. detail. DRL adopts,. and explain. namely. the criteria, the evaluation, and the issue space. Figure 4.4. Model. how. they represent. the argument, the alternative,. shows. the regions of a decision. graph schema which represent the different spaces.. The Argument Space. An argument. is. represented in. DRL. A. as a set of related Claims.. other people might call facts, assumptions, statements, or rules. distinctions,. which. is. sometimes arbitrary and. attribute Plausibility that indicates. advantage of not imposing a. set. how much. it is. desirable to. make. confidence. Instead of. make, a. we have. DRL. making. these. Claim has. in the claim.. the. This has the. of predetermined categories on the user, and avoiding the. ambiguity resulting from the disagreement. When. difficult to. Claim subsumes what. among people on what. the distinction say,. between. facts. facts or. assumptions. and assumptions,. we. are.. can do. so simply by specializing a claim or by using nominal categories like "fact" and. "assumption" based on the Plausibility attribute values. do so post facto or dynamically by using. mapping between. this. measure and. the. a. in different. Claims. In. fact,. we can. numeric measure as the plausibility value, and. measure based on nominal categories. "assumptions.". 46. like "facts " or.

(61) ssue Space. O. Criteria Space. <:^^-u. a atbgOAi ot. fubgoti. d. XD. Alternative. Space. Argument Space. Os. adseva. is-«-re>uit-oi'. c_z> Figure 4.4. The. constructs of. DRL grouped. by the kinds of decision rationales. that they. represent.. A. Claim can be Supported, Denied, or Presupposed by another Claim. For example,. Claim, [Difficult. to use), is. lower. DRL. Claim. level].. 1. to. All. 2].. a. Supports. work. at a. For example, when we. link. the claim, [Requires application to. relations are special types of Claims.. Claim 2 through. supports Claim. supponed by. relation,. we. are. making. the claim, [Claim. Supports, Denies, Achieves,. Is. A Subgoal. users can suppon, deny, or qualify them.. Reduce dvt. cost]. cost]) is a claim. 1. Likewise, an Achieves relation from an Alternative object to a Goal. object represents the claim that the alternative achieves the goal. Hence, any like. the. (shown. which. is. Of,. is. DRL relation,. a Claim, and can be argued about;. For example, [Interaction manager achieves. as the achieves link. denied by [Expensive. 47. between [IM] and the goal [Reduce dvpt to build].. That [Expensive. to build] denies.

(62) the achieves claim. is itself. a claim, which. is. in. tum denied by. the claim [This. is. one time. cost].. The Alternative Space. Alternative. is. the fundamental unit of the Alternative. Space and represents an option being. considered, e.g. "Toolbox" and "Interaction Manager." Alternatives are. boxes. in the left. bottom of the. Currently, in. figure.. other through only the generic relation. Is. IM]. is. a special case of the alternative, [IM].. imponant Version. to represent. of.. among. alternatives are related to each. Thus,. we can. may. There. Although these relations may be added. to. DRL if the. we can do. We. Is. linking. them. it. application.. deny the. Is. it. represents the. really a specialized version of another. is. is. A Kind of. The. Presupposes. He. A Kind. For example, the user. really not a special kind of [IM]. can express (9/". relations can be also qualified by. relation.. violates the fundamental characteristic of [IM],. making. same way. can argue about whether an alternative should be an. or Is. to another claim via a. argue that [Extensible IM]. manager and. task requires them, they. so by creating Claims that deny or support the appropriate relations,. A Good Alternative for. such as. Next. DRL.. alternative at all or whether an alternative alternative;. say that [Extensible. be other relations that are. represents the arguments about the alternative space the. arguments about the goal space.. to. Of.. as rounded. alternatives, such as Elaborates, Simplifies, or Is the. are not in the core vocabulary of. DRL. A Kind. DRL,. shown. relation. this. namely the. 48. because the extensibility. strict. separation of. argument by adding a claim. between the two. may want. alternatives.. window. to that effect. and.

(63) The. In. Criteria Space. DRL,. by Goals.. criteria are represented. DRL. "Criterion" because for each criterion,. we can always. the goal of achieving the criterion.. We. possible. among. A. define a corresponding goal, namely. want. to. convey the richer relationship. these goals than what the term criteria usually conveys.. Goal Is A Subgoal o/ another Goal the second.. rather. uses the term "Goal" rather than. set. if. achieving the. of subgoals can be related. first. Goal. For example, a. facilitates the. among themselves. in various. achievement of ways; they can. These. be mutually exclusive, independent of each other, or panially overlapping. relationships are represented by creating a. members; the. relations. among. Group. object and specifying these Goals to be. Goals are specified. these. in the. "Member. its. Relations" property. of the Group.. Decision Problem represents the goal of choosing the best altemarive. All the other goals for a decision are subgoals of the decision. problem. means. our example, the interpretation of the decision. to. choose the best. alternative.. In. problem [Which application interface for application interface that [Is. ponable],. is. is. best for our. WING?]. is. [Is. what. it. the desired state of having chosen an. window manager, and. the desired state of having. window manager ponable. Hence,. in the sense that they elaborate. the interpretanon of the goal,. chosen an application interface that makes our. portable]. is. a subgoal of the decision. problem. in. this interpretanon.. The arguments about alternative space. relation,. we can. the criteria space are represented in the. Because the. Is. A Subgoal of relation. argue about whether a goal. is. is. same way. a Claim, as. desirable or whether. as those about the. is. it. any other. contributes to. achieving another goal by arguing about this relational claim. For example,. about whether [Support. many UI. styles]. should be a goal. 49. at all. DRL. we can. argue. by producing claims.