NUDGE, A Knowledge-Based Scheduling Program

To implement the N U D G E program, a knowledge representat ion language — FRL-0 — based on a few powerful generalizations of the t r a d i t i o n a l p r o p e r t y l i s t representa t ion has been developed. T h e N U D G E knowledge base defined in FRL-0 consists of a h ierarchica l set o f concepts t ha t p rov ide generic deser ip t ions of the t y p i c a l ac t iv i t ies , agents, plans and purposes of the domain to be scheduled. C u r r e n t l y , t h i s domain is the management and c o o r d i n a t i o n of personnel engaged in a group project. N U D G E const i tutes an experiment in knowledge-based, ra ther than power-based AI programs. It also provides an example of an in te l l igent suppor t system, in which an AI program serves as an aid to a decision maker. Final ly, NUDGE has served an exper imental vehicle fo r test ing advanced representation techniques. 1. I N T R O D U C T I O N A classic issue in AI is the knowledge versus power controversy [ M i n s k y & Paper t 74] . T h e knowledge pos i t ion advocates t ha t intel l igence arises mainly f rom the use of a large store of specific knowledge, whi le the power theory argues for a small col lect ion of general reasoning mechanisms. This paper reports on an experiment in which a knowledge-based program N U D G E has been implemented fo r the schedul ing domain, a domain in which power-based programs have long been the dominant paradigm. T rad i t i ona l l y , scheduling programs apply simple but power fu l decision analysis techniques to f inding the opt imal schedule under a wel l -def ined set of constraints. The performance of N U D G E conf i rms t ha t f o r wel l -def ined, formal si tuations, the t rad i t ional power-based approach is app rop r i a t e . B u t fo r the prob lem of d e f i n i n g these f o r m a l s i t u a t i o n s when g i v e n on ly i n f o r m a l s p e c i f i c a t i o n s , a knowledge-based approach is necessary. By an informal speci f icat ion, wc mean a scheduling request that is potential ly incomplete, possibly incons is ten t and q u a l i t a t i v e . (See Balzer [ 74 ] fo r an analys is of i n fo rma l program specifications.) Thus, the N U D G E program accepts i n f o r m a l requests and produces a calendar con ta i n i ng possib le c o n f l i c t s and an associated set of s t ra teg ies fo r reso lv ing those conf l ic ts . A domain-independent search a lgor i thm B A R G A I N then resolves these conf l ic ts by t radi t ional decision analysis techniques. N U D G E uses a broad data base of knowledge to expand and debug in formal scheduling requests. The database is used to supply missing detai ls, resolve inconsistencies, determine available opt ions, not ice necessary prerequisites and plan for expected outcomes. To manage th is large store of knowledge, a representation language -F R L 0 — has been implemented. F R L 0 extends the t r a d i t i o n a l a t t r i b u t e / v a l u e descript ion of properties by al lowing propert ies to be described by comments, abstractions, defaul ts, constraints, ind i rect pointers f r o m other properties, and attached procedures. These are not new representation techniques. Abst ract ion, for example, was discussed in Qui l l ian [68], and attached procedures have become a common property of AI languages since P L A N N E R [Hew i t t 69]. However, the strengths and weaknesses of these representat ion techniques and the i r potential interactions is s t i l l not well understood. For th is reason, we have chosen not to include as many representat ion capabi l i t ies as are cur ren t l y being implemented in K R L [Bobrow & W i n o g r a d 7 6 ] and O W L [ M a r t i n 7 7 ] . W c v iew F R L 0 a s a n e x p e r i m e n t a l m e d i u m to s t u d y t he u t i l i t y o f a few s p e c i f i c capabi l i t ies and the i r interactions. Because a knowledge-based approach requires a large store of specif ic data, it was necessary to choose a par t icu lar domain to car ry ou t ou r experiment. Our cr i ter ion was to select a realm in which schedul ing requests are typical ly in formal . Th is c r i te r ion ru led out such schedul ing problems as those of an assembly line. (See Tong r [63 ] fo r an AI t reatment of this problem.) Instead, we selected of f ice schedul ing; in par t icu lar , assisting a manager in scheduling his team. T h i s e n v i r o n m e n t inc ludes schedul ing meet ings, m o n i t o r i n g the progress of subgoals assigned to team members, a ler t ing the manager to deadlines, and rea l t ime rescheduling. I n p r o v i d i n g N U D G E w i t h the knowledge necessary f o r these funct ions, our research serves a th i rd purpose beyond (1) exp lor ing the re la t ion between knowledge-based and power-based schedul ing and (2) exe rc i s ing va r ious representa t ion s t ra teg ies . I t p rov ides ins ight in to the categories of knowledge that are necessary for o f f ice schedul ing (independent of thei r representation). N U D G E contains a h ierarchy fo r act iv i t ies involving informat ion t ransfer, for people in var ious roles related to this transfer, for the plans governing these t r a n s f e r s , and f o r t h e assoc ia ted demands on t i m e , space and personnel . T h e h i e ra r chy is on the average f i ve levels deep and includes approximately 100 objects, each described by a generalized proper ty l is t called a frame. An abridged version of th is h ierarchy appears below, w i t h specialization indicated by nesting. Knowledge 257 ■4: Goldste in T h e t e r m " f r a m e " a s used i n F R L 0 w a s i n s p i r e d b y M i n s k y ' s [7 .5 ] deve lopment o f f r ame theory . Frame theory contends t ha t ( ! ) intel l igence arises f rom the application of large amounts of h igh ly s p e c i f i c k n o w l e d g e , a t opposed to a few g e n e r a l i n f e r c n o n g mechanisms, and (2) th is is accomplished th rough the use of a l ib rary of frames, packets of knowledge that provide descriptions of typ ical ob jec ts and events . These descr ip t ions conta in bo th an abs t rac t template prov id ing a skeleton for describing any instance and a set of de fau l ts fo r typical members of the class. The defaul ts allow the in fo rmat ion system to supply missing detai l , maintain expectations, and notice anomalies. We have yet to inves t iga te w i t h equal care re la ted a rea* of knowledge not s t r i c t l y involved in scheduling the in format ion f low between members of a research team — this includes space al locat ion, budge t ing , and t ravel scheduling. The last of these is the focus of the GUS system developed by Robrow ct al [76]. Gus IS is a f r amebased dialog system that addresses many of the issues raised here. T h e ma jo r d i f f e rence is t h a t GUS focusses on the use of f r a m e rep resen ta t i ons to suppo r t d ia log comprehension wh i le N U D G E focusses on thei r use to support sophisticated scheduling. F inal ly , a l though developed for the off ice scheduling domain, the NUDGE knowledge base does have a broader appl icabi l i ty. The use of a h ierarchy of successively more abstract concepts allows h igh ly specif ic in fo rmat ion to be clearly segregated f rom more general facts about the wor ld . Fur thermore, some of the concept hierarchies such as those fo r t ime and place apply in toto to many other applications.