sign in sign up
<<
Home , Metaknowledge

Domain : Enabling User-Centric Planning

Karen L. Myers Artificial Intelligence Center SRI International 333 Ravenswood Ave., Menlo Park, CA 94025 [email protected]

Abstract planning base could completely capture the sub- tleties of a particular domain, users often want to participate Substantial improvements must be made in the usabil- in the plan development process to some extent, in order to ity of AI planning technology in order for it to achieve ensure that they understand both the structure of a plan and widespread adoption. In particular, approaches are re- quired that better serve the needs of users, who gen- the process by which planning decisions were made. erally want to play a central role in the plan develop- For these reasons, the future of automated planning lies in ment process. This user-centric perspective introduces the development of user-centric planning tools that can as- new requirements for the design of planning systems. sist human decision-makers rather than replace them. These In this paper, we argue that increased usability requires tools should enable a user to understand the complexity of a new representational layer that captures metatheo- the underlying problem, as well as providing guidance in retic properties of the planning application. A domain the determination of a solution that is well-suited to their would provide an abstracted characteriza- specific needs. tion of the planning domain that highlights key seman- tic differences among planning elements. This paper The main thesis of this paper is that effective interaction presents a candidate model for a domain metatheory, between users and planning technology requires an augmen- as well as an instantiation of that model for a simple tation of standard planning models to include an explicit travel domain. In addition, it describes three capabil- domain metatheory. A standard planning domain is mod- ities for user-centric planning that are enabled by this eled in terms of three types of elements: individuals cor- model, namely user directability of planning, genera- responding to real or abstract objects in the domain, rela- tion of qualitatively different plans, and plan summa- tions that describe characteristics of the world and individual rization. world states, and operators that describe ways to achieve ob- jectives. The domain metatheory would capture high-level Introduction semantic attributes of planning elements, thus providing a rich vocabulary for describing characteristics of solutions Artificial intelligence (AI) planning technology provides and problem-solving processes. As argued in this paper, powerful tools for solving problems that require the coor- the availability of such a language would greatly facilitate dination of actions in the pursuit of specified goals. The AI a broad range of user-planner interactions. community has produced several planning systems whose Different types of metatheories may suit different pur- demonstrations on realistic problems attest to the value of poses. This paper presents a candidate model for a plan- automated planning techniques. Nevertheless, there has ning metatheory, along with an instantiation of that model been limited success in transitioning this technology to user for a simple travel domain (Mayer 1997). The paper de- communities. A major reason for the lack of technology scribes how this metatheory was used to support two critical transfer lies with the difficulty of using planning systems. user-centric planning capabilities: user directability of the AI planners have traditionally been designed to operate as planning process and the generation of qualitatively differ- “black boxes”: they take a description of a domain and a set ent plans. In addition, a proposal is put forth for using the of goals and automatically synthesize a plan for achieving metatheory in support of plan summarization. the goals. This design explicitly limits the amount of influ- ence that a user can have on the generated plans. In many domains, users are reluctant to relinquish full A Domain Metatheory control of the planning process. This reluctance stems from Overview several factors. One is a that human creativity and Our candidate domain metatheory is built around three main experience are essential for effective planning in complex constructs: roles, features,andmeasures. domains. Transparency is another: even in the event that a A feature designates an attribute of interest for an oper-

Copyright c 2000, American Association for Artificial Intelli- ator that distinguishes it from other operators that could be gence (www.aaai.org). All rights reserved. applied to the same task. For example, among operators that g can be used to refine tasks of moving from location X to lo- Vacation-Scope = fOverseas National Regional cation Y, there can be some that involve travel by air, land, AFFORDABILITY: (Overseas National Regional) (Overseas National Regional) or water; each of these media could be modeled as a fea- TIME-EFFICIENCY:

ture. Because there can be multiple operators that apply to g Accommodation = fHotel Motel Camp a particular task, features provide a way of abstracting from COMFORT: (Camp Motel Hotel) the details of an operator up to distinguishing attributes that AFFORDABILITY: (Hotel Motel Camp)

might be of interest to users. Note that features differ from g operator preconditions in that they do not directly restrict Transport-Media = fAir Land Water use of operators by the planner. AFFORDABILITY: (Water Air Land)

Related features are grouped into feature categories.For TIME-EFFICIENCY: (Water Land Air) g

example, the features fAir Land Water mentioned above g define a Transport-Media category. Feature categories Land-Transport-Mode = fAuto Bus Shuttle Taxi Train Limo AFFORDABILITY: (Limo Train Auto Taxi Shuttle Bus) themselves can have interesting properties. Just as planning TIME-EFFICIENCY: (Bus Shuttle Auto Limo Taxi Train) operators reflect a hierarchical structure, features and fea- COMFORT: (Bus Shuttle Taxi Train Auto Limo)

ture categories can be organized hierarchically. Certain cat- g egories may be mutually exclusive in that at most one feature Transit-Ownership = fPublic Private from the category can be assigned to any given operator; this COMFORT: (Public Private)

is the case for the feature category Transit-Ownership con- AFFORDABILITY: (Private Public) g taining the elements fPublic Private . Other categories TIME-EFFICIENCY: (Public Private)

may support overlapping features; for example, there may g be an operator that involves both Air and Land travel. Transit-Capacity = fSolo Shared A role corresponds to a capacity in which a domain COMFORT: (Shared Solo) AFFORDABILITY: (Solo Shared) object is to be used within an operator. Roles map to individual variables within a planning operator. For in- stance, an air transportation operator could have variables Figure 1: Feature Categories and Associated Measures from location.1 and location.2 that fill the roles of the Travel Domain Origin and Destination,aswellasavariableairline.1 that fills the role of Carrier. A comparable sea transporta- tion operator may have these same roles, although with the ranks higher than Ritz with respect to AFFORDABILITY. planning variable cruise-ship.1 for the role Carrier. We define the domain of a measure to be the set of (par- Feature categories can have associated measures.A tially) ordered values employed by the measure. For mea- measure corresponds to an ordering (possibly partial) of sures defined over feature categories, the domain is the set features within the category with respect to some desig- of features that comprise the feature category. For measures

nated criteria. For example, consider the feature category defined over instances, the domain is the set of measure val- g Transit-Ownership with features fPublic Private .For ues that can be assigned to instances. the measure COMFORT, the feature Private would rank higher than Public; for the measure AFFORDABILITY,the order would be reversed. Discussion Figure 1 presents an excerpt from the metatheory for the There is no ‘correct’ formulation of a metatheory: as with travel domain that shows sample feature categories and as- the underlying planning domain, its design involves an ex- sociated measures. Each block defines a feature category, plicit modeling process. Different potential users may care with the first line listing the name of the feature category about different metatheoretic properties: for example, af- followed by its constituent features. The remaining lines de- fordability may be an issue when designing a system to be clare a measure associated with that feature category, and used for students, but not for business travelers. provide the ranking of the features for that measure and cate- gory. For simplicity, we show only measures that completely The value of the domain metatheory lies with its provi- order the features; however, partial orders are possible. sion of a semantically-grounded abstraction of the under- Just as measures can be employed to rank features (and lying planning domain that enables meaningful, high-level hence operators with those features), they can also be em- descriptions of plans and planning processes. As part of the ployed to rank instances. For measures on instances, an abstraction process, the metatheory provides semantic link- ordered set of measure values is defined. For each mea- age among different elements within a planning domain: sure, a given individual can (optionally) be assigned one

 Operators may share a common feature. of these values, thus inducing a partial order over in-

stances. In the travel domain, for example, the measure  Operators and instances may be related by a measure. AFFORDABILITY has the values (Extravagant Expensive Moderate Inexpensive Cheap) in increasing order from left As shown below, this linkage enables concise descriptions of to right. The individual Ritz of class Hotel has the plan properties, both on the part of a user seeking to direct a AFFORDABILITY value Extravagant, while the individual planning system, and a system seeking to summarize plans Motel6 of class Motel has the value Cheap, thus Motel6 or planning decisions for a user. User Directability: Planning Advice renderings of advice to aid understandability, but it is easy to One natural way to support the use of automated planning map to our structured model. For the first example, the con- technology without ceding full control is to allow users to text activity is defined as tasks with feature Vacation,and direct the operations of the underlying planning technology with role Location filled by Scotland. The advice dictates by specifying desired plan attributes. For example, a traveler that the filler for the role Accommodation be an object that could express preferences for a particular trip (e.g., modes belongs to the class 3-star-hotel and have ensuite facilities of transportation for various legs, specific airlines to use, listed as an attribute. accommodation requirements, and restrictions on costs for various aspects of the trip), with an automated planner con- Method Advice Method advice imposes restrictions on structing a solution that seeks to maximize satisfaction of the approaches that can be used in solving a goal those preferences. or class of goals. It is characterized by the tem- Our advisable planning framework embodies this model plate: of user directability (Myers 1996; 1999; 2000). It enables for Thus, method advice users to provide advice to an automated planning system consists of context and advised activities, along with a po- in order to influence the content of the solutions that are larity expressing prescription or proscription. For example: produced. Advice consists of session-specific recommenda- tions on how tasks are to be accomplished, in terms of spe- Find a package bike tour starting in Athens for the cific approaches to pursue and entities to employ. Advice is vacation in Greece. specified in a high-level language designed to be natural and Don’t fly between cities less than 200 miles apart. intuitive for users, and then operationalized into constraints The first piece of method advice declares that the approach that direct the underlying planning technology during plan used for a particular portion of the trip should have cer- construction. The advice language builds on the features and tain features (i.e., Bike, Package) and role constraints (i.e., roles from the domain metatheory, along with the underlying Start-Location is Athens). The second specifies restric- language used to represent a given domain.1 tions on the approach to be taken for solving a class of trans- port goals. Advice Specification Advice comes in two forms: role and method. Role advice Advice Enforcement constrains the use of domain entities in solving tasks, while Models for satisfaction of advice by a hierarchical task net- method advice further constrains the type of approach used. work (HTN) plan are provided in (Myers 1996) along with Both types are formulated in terms of role-fills and activities. algorithms for strict enforcement. Algorithms for relaxed Activities represent abstract tasks relative to the underlying satisfaction can be found in (Myers 2000). The basic ap- planning domain, and are defined in terms of features and proach involves adding advice constraints that focus the roles from the domain metatheory. A given activity can have planner on choices (for operators and variable instantia- multiple features; for example, an activity corresponding to tions) whose metatheoretic properties are compatible with a bike tour could have the features Vacation, Bike,andIn- the user-specified advice. expensive. Role-fills are specifications of objects to be used in certain roles; they may name explicit individuals, or con- Generating Qualitatively Distinct Plans sist simply of a set of required and prohibited attributes. Many real-world applications have solution-rich search spaces. Air campaign planning (Thaler & Shlapak 1995; Role Advice Role advice either prescribes or restricts Lee & Wilkins 1996) and travel planning (Linden, Hanks, & the use of domain entities for filling certain capacities Lesh 1997)) provide two examples. For these applications, in the plan. Role advice is characterized by the tem- it is not difficult to find a solution; rather, the challenge for plate: in human planners is to understand the range of available op- for In general, role advice tions in order to ensure informed selection of a solution. One consists of one or more role-fill specifications, a context ac- means by which to help users with this task is to provide a tivity,andapolarity indicating whether the advice is pre- set of qualitatively distinct plans distributed throughout the scribing or prohibiting the role-fill. The following directives overall solution space, thus providing exemplars of possible provide examples of role advice: approaches. Stay in 3-star ensuite hotels while vacationing in Current automated planning tools can readily generate Scotland. different plans, for example through repeated runs with ran- Layovers longer than 90 minutes are unaccept- domized choices at decision points. The differences among able for domestic flights. such plans, however, are difficult to extract and not neces- The first directive imposes requirements on accommoda- sarily semantically meaningful. Furthermore, different users tions during vacations in a given area. The second prohibits will have different notions of what constitutes ‘meaningful’ flights with long layovers. Here, we use natural language differences. For example, a budget traveler might like to see options with a range of costs while the business trav- 1Measures could be used as the basis for more advanced forms eler might like to see options that maximize accumulation of advice but have not been explored to date. of frequent flier points. Ideally, a system for generating qualitatively different plans would allow the user to specify Plan Summarization and Comparison dimensions along which he or she would like to see varia- The usability of automated planning technology would be tion. Recent work on mixed-initiative, interactive, and ad- enriched greatly by an automated plan summarization capa- visable planning enables users to drive the process of gener- bility that could convey the key aspects of generated plans. ating qualitatively different plans (Ferguson & Allen 1998; Meaningful plan summaries would enable humans to feel Tate, Dalton, & Levine 1998; Myers 1996). With these more comfortable with delegation of planning tasks to an frameworks, however, the user must be involved extensively automated system, knowing that the essence of the solution in an ongoing role to articulate desired differences and to will be concisely and completely presented to them. Simi- manage the space of options. larly, the ability to present comparisons of plans that high- Our work on metatheoretic biasing leverages the domain light key differences would help users navigate more effec- metatheory described above to enable generation of qualita- tively through large solution spaces to identify plans that are tively different plans (Myers & Lee 1999). In particular, the well-suited to their needs. approach capitalizes on the structure inherent to measures to There has been limited work to date on summarization create biases that focus the planner on solutions with certain and comparison of plans, with most efforts focused on meth- attributes. Biases are selected in a manner designed to pro- ods that are grounded in syntactic characteristics (Young duce multiple solutions from different regions of the overall 1999). In contrast, the domain metatheory provides the po- plan space. tential to abstract from the details of plan structures to con- To generate n plans, the method partitions the domains of cise summarizations of key decisions within plans, and to selected metatheory measures into n intervals. Each inter- important differences among plans. val from a measure is assigned to one of nbiassets, with We envision an approach that employs a suite of tech- the different sets being used to generate different plans.2 niques to identify patterns or exceptions relative to metathe- For example, to generate two plans using the measures oretic structures, along the lines of the following. AFFORDABILITY and COMFORT, the domains of these mea- Role/Feature Abstraction Role abstraction involves uni- sures would be split into two subsets, with one correspond- versal quantification over values selected to fill designated ing to a high valuation and the other a low valuation. Dif- roles: ferent algorithms can be used to assign the n intervals from United was chosen as Carrier for all air transportation. each measure to the n bias sets: one might establish a first Similarly, feature abstraction involves universal quantifi- bias set with high affordability and low comfort biases and cation over approaches selected to achieve similar goals: the second with low affordability and high comfort biases. Public-transport was used throughout for transit to and Biases are enforced during planning in a heuristic man- from airports. ner: rather than imposing hard constraints, choices available to the planner (namely, operator selection and instance se- Measure Relativization A generalization of role/feature lection) are ordered to reflect their distance (according to abstraction, this method would enable summarization rel- the metatheory measures) from the stated biases. Because ative to measures: the enforcement of biases prioritizes choices rather than fil- Accommodations were chosen that ranked high on af- tering them, it does not restrict the set of plans that could be fordability. produced. As such, the biases can be viewed as relaxable Role/Feature Differencing This method would enable constraints on plan generation. highlighting of key differences between two plans by By capitalizing on the semantic structure of domain noting variations in the way that roles are filled or that metatheory measures, the metatheoretic biasing technique choices are made among competing operators with differ- provides a low-cost mechanism for the generation of plans ent features: with meaningful semantic differences. In particular, it en- United was chosen as the Carrier for air travel in Plan-1 ables the generation of plans that are qualitatively different while Delta was chosen in Plan-2. by design, rather than relying on random search through the Plan-1 uses Hotels for accommodation while Plan-2 in- syntactic plan space. The experimental results in (Myers & volves Camping. Lee 1999) validate the effectiveness of the method for re- liably generating a range of plans with meaningful seman- These techniques would be used in conjunction with a tic differences. Additionally, users can optionally direct the summarization profile specified by a user that indicates those planner into desired subregions of the overall plan space by aspects of the metatheory that interests him or her most. designating specific measures and subintervals of those mea- sures that should be used for bias generation. For exam- Knowledge Costs ple, users could indicate that they want to see plans within Fully-automated planning systems are brittle in that they re- a range of cost and time values, while insisting on traveling quire complete and correct formalizations of the domain in by airplane (rather than train, boat, or car). order to function. Small omissions or errors can result in the inability of systems to yield any solutions. Providing comprehensive domain is time-consuming and 2This overview is necessarily simplified; (Myers & Lee 1999) expensive, and represents a significant investment for each provides a comprehensive description of the approach. new application. In rich application domains, those models will need to grow and evolve over time, further exacerbating Acknowledgments This research has been supported the knowledge acquisition problem. by DARPA Contract F30602-95-C-0259 as part of the Given the inherent cost and difficulties in building and ARPA/Rome Laboratory Planning Initiative. maintaining complex knowledge bases, is it practical to advocate the inclusion of metaknowledge in planning do- References mains? Cost notwithstanding, we believe that this informa- Ferguson, G., and Allen, J. 1998. TRIPS: Towards a tion is essential to providing effective usability of advanced mixed-initiative planning assistant. In Proceedings of the planning tools. Furthermore, while development of a do- AIPS Workshop on Interactive and CollaborativePlanning. main metatheory does increase the amount of knowledge Gil, Y., and Swartout, B. 1994. EXPECT: A reflective ar- acquisition required to develop a planning application, there chitecture for knowledge acquisition. In Burstein, M. H., are several factors that mitigate the overall cost. ed., ARPA/Rome Laboratory Planning and Scheduling Ini- Reduced Sensitivity A domain metatheory is much less tiative Workshop Proceedings. San Mateo, CA: Morgan sensitive to errors than are the base-level planning theories Kaufmann. since the content does not alter the set of legal solutions. Lee, T. J., and Wilkins, D. E. 1996. Using SIPE-2 to in- For this reason, metatheory inaccuracies may lead to unex- tegrate planning for military air campaigns. IEEE Expert pected solutions or incompleteness when exploited by the 11(6):11–12. algorithms for advisability or generation of qualitatively dis- tinct plans, but will not impact the basic planning process. Linden, G.; Hanks, S.; and Lesh, N. 1997. Interac- tive assessment of user preference models: The Automated For example, within our candidate metatheory, missing Travel Assistant. In Proceedings of the Sixth International features could result in an inability to find solutions that Conference on User Modeling. maximize satisfaction of stated user; however, some solu- Mayer, M. A. 1997. A natural language interface for the tion will still be returned whenever the underlying problem Advisable Planner. Senior Honors Thesis, Stanford Uni- is solvable. When generating qualitatively different plans, versity. the distinction between certain semantic differences may be lost, resulting only in a missed opportunity to show the user Myers, K. L., and Lee, T. J. 1999. Generating qualita- an interesting dimension of plan variability. tively different plans through metatheoretic biases. In Pro- ceedings of the Sixteenth National Conference on Artificial User Initiative The adoption of metatheories will enable Intelligence. AAAI Press. greatly increased user involvement in the planning process. This in turn will lessen the requirements for correctness and Myers, K. L. 1996. Strategic advice for hierarchical plan- comprehensiveness imposed on the underlying knowledge ners. In Aiello, L. C.; Doyle, J.; and Shapiro, S. C., bases, since the user could be expected to share responsi- eds., Principles of Knowledge Representation and Reason- bility for both planning knowledge and plan validity. For ing: Proceedings of the Fifth International Conference (KR example, in cases where a planner is unable to find a solu- ’96), 112–123. Morgan Kaufmann Publishers. tion, interaction with the user may help to identify problems Myers, K. L. 1999. User Guide for the Advisable Planner. with the knowledge base that incorrectly eliminated poten- Artificial Intelligence Center, SRI International, Menlo tial branches in the search space. Park, CA. Version 2.3. Ease of Formulation We believe that a good domain Myers, K. L. 2000. Planning with conflicting advice. In metatheory should be a natural by-product of a principled Proceedings of the Fifth International Conference on AI approach to knowledge acquisition and modeling for plan- Planning Systems. ning knowledge bases. For example, when defining multiple Tate, A.; Dalton, J.; and Levine, J. 1998. Generation of operators that overlap in their applicability, users could be multiple qualitatively different plans. In Proceedings of the required to assign features to indicate how those operators Fourth International Conference on AI Planning Systems. differ at a semantic level. The metatheory elicitation pro- cess would be facilitated by the use of structured knowledge Thaler, D. E., and Shlapak, D. A. 1995. Perspectives on acquisition tools such as EXPECT (Gil & Swartout 1994). theater air campaign planning. Technical report, Rand Cor- poration. Young, R. M. 1999. Cooperative plan identification: Con- structing concise and effective plan descriptions. In Pro- Conclusions ceedings of the Sixteenth National Conference on Artificial Intelligence. AAAI Press. This paper has argued that an explicit domain metatheory will play a critical part in effective user-centric planning sys- tems. This thesis has been supported by the presentation of a specific model for a metatheory, along with a description of how it was used to support user advisability of planning systems and the generation of qualitatively distinct plans. A proposal for plan summarization methods shows promise for future additional applications of metatheories.