The Restricted Language Architecture of a Hybrid Representation System

Hybrid architectures have been used in several recent knowledge representation systems. This paper explores some distinctions between various hybrid representation architectures, focusing in particular on systems built around restricted representation languages This restricted language architecture is illustrated by describing KL-TWO, a hybrid reasoner based on the restricted representation facility RUP. The bulk of this paper discusses KLTWO, its subcomponents, and the techniques used to interface them. Many recen t knowledge representa t ion systems have been based on h y b r i d a rch i tec tu res . These systems are hybr ids in t h a t they do not at tempt the en t i re knowledge rep resen ta t i on task w i th a single i n fe ren t i a l component. Ins tead, they combine several reasoners i n to a complex whole. The advantages to be gained by th i s h y b r i d approach vary from system to system, bu t o f ten inc lude increases in the system's computa t iona l e f f ic iency, the coverage of i ts rep resen ta t ion language, or the ease of expressing knowledge w i th the system. The top ic of th i s paper is a new l og i c -based hyb r i d r ep resen ta t i on system. This system, cal led KL-TWO, is one wh ich I have been developing w i th my colleagues over the past two years. Al though KL-TWO has fea tu res in common w i th other hyb r i d reasoners, i t d i f fe rs ve ry s ign i f i can t l y from other systems in the way the common fea tures are used. The purpose of th i s paper is t hus twofo ld , to describe the features t h a t KL-TWO possesses, and to stress the di f ferences in approach between KL-TWO and other hyb r i d reasoners. Hybrid Representation Architectures The h y b r i d rep resen ta t ion scheme most fami l iar to the genera l AI audience is t ha t embodied by such systems as KRYPTON (Brachman, Fikes, & Levesque, 1983) or the t h e o r e m p r o v i n g system of Schubert and his col leagues (Schuber t , Papalaskar is , & Taugher, 1983). These systems s t a r t w i th a theorem prover for f i r s t o r d e r logic (FOL), and augment i t w i th spec ia l -pu rpose in fe rence systems. These spec ia l -pu rpose components are designed to per form ef f ic ient ly ce r ta in forms of reason ing t h a t would be computat iona l ly expensive fo r the theorem prover to per form by i tsel f . In essence, f o r t h e i r r e s t r i c t e d domains, these augmentat ions bypass the normal proof procedure of the theorem p rover , and cu t out some of i t s combinator ia l search. For example, KRYPTON augments St ickel 's connec t ion graph theorem prover w i t h a device fo r te rm ino log ica l reasoning. Schuber t and his colleagues added to t h e i r theorem prover components to reason about types, p a r t w h o l e re la t ionsh ips , tempora l o rde r ing , and co lor . In th is paper, I wi l l re fer to systems such as these as theorem prover-based hybrid reasoners. One impor tan t cha rac te r i s t i c of these theorem p r o v e r b a s e d hybr ids i s t h a t the h y b r i d a r c h i t e c t u r e may enhance the overa l l ef f ic iency of the system, b u t wi l l not change i ts overa l l expressiveness. Any th ing which can be represen ted in the spec ia l pu rpose augmentat ions could also be represented in the unadorned theorem prover (a lbei t less conven ien t ly or e f f ic ient ly ) . The augmentat ions don ' t ex tend the expressive power of the theorem prover as much as they c i rcumvent i t s computa t iona l inef f ic ienc ies. KL-TWO has a subs tan t ia l l y d i f fe ren t a r c h i t e c t u r e f rom t h a t of theorem p r o v e r b a s e d systems. This a l te rna t i ve a r c h i t e c t u r e was f i r s t suggested by McAllester (1980, 1982). He argued t h a t one need no t base a rep resen ta t ion system on a f u l l in ference system for the pred ica te ca lcu lus. Indeed, given t h a t f u l l f i r s t o r d e r logic is only semi -dec idab le , and t h a t even the best t h e o r e m p r o v e r s are prone to t ime -consum ing search, the re is good reason to consider a l t e rna t i ve rep resen ta t i on schemes. McAllester proposed bu i l d ing knowledge rep resen ta t i on systems a round a f a i r l y simple r e s t r i c t e d device, his Reasoning U t i l i t y Package (RUP for shor t ) . RUP provides only a subset of t he i n fe ren t i a l power of a f i r s t o r d e r theorem p rover , ( rough ly the p ropos i t i ona l subset) b u t is computa t iona l ly much more ef f ic ient . To res to re some of the missing rep resen ta t i ona l power, one extends RUP w i th spec ia l pu rpose reasoners t a i l o red to one's app l ica t ion . Other r e s t r i c t e d languages have been proposed fo r th is k i nd of h y b r i d system ( for example, the re levance vers ions of p ropos i t i ona l logic and FOL proposed by Levesque (1984) and Pa te l -Schne ide r (1985)). I use t h e te rm restricted language hybrid reasoners to re fe r to hybr ids bu i l t a round r e s t r i c t e d systems such as those of McAllester, Levesque, or Pa te l -Schne ide r . The advantages of the r e s t r i c t e d language approach become apparent i f the subcomponents t h a t ex tend the c e n t r a l reasoner are themselves r e s t r i c t e d b u t e f f ic ient . The resu l t i ng system wi l l no t have the f u l l r ep resen ta t i ona l expressiveness of a f i r s t o r d e r language, bu t in exchange i t w i l l have the p o t e n t i a l f o r reasoning e f f ic ien t ly . This approach has a ve ry a t t r a c t i v e promise: i f one can make do w i t h t h e expressive l im i ta t ions of a r e s t r i c t e d language h y b r i d , one may have a very e f f i c ien t system indeed. This restorer, wot tupporttd by the Advonct Rt t torch Projact Agency of th t Dtportment of D t f t n t t . Under Contract No. NO0014-77-C-6378.