Methodological simplicity in expert system construction: the case of judgements and reasoned assumptions
Jon Doyle · Morgan Kaufmann Publishers Inc. eBooks · 1989
P r o b a b i l i s t i c rules a n d their v a r i a n t s have recent ly s u p p o r t e d severa l success fu l a p p l i c a t i o n s of e x p e r t s y s t e m s , in s p i t e of the difficulty of c o m m i t t i n i n f o r m a n t s to p a r t i c u l a r condi t iona l p robab i l i t i e s or cer ta in ty f a c t o r s , a n d in sp i t e of t h e exper imenta l ly obse rved insens i t iv i ty of s y s t e m p e r f o r m a n c e to p e r t u r b a t i o n s of t h e chosen va lues . Here we survey recent d e v e l o p m e n t s concerning r e a s o n e d ass u m p t i o n s which offer hope for avo id ing the p r a c t i c a l e lus iveness of p robab i l i s t i c rules while re ta in ing theoret ica l power, for b a s i n s y s t e m s on the i n f o r m a t i o n unhes i t a t ing ly a i n e d f rom exper t in formant s , and r e c o n s t r u c t i n t h e enta i led degrees of belief l a ter . I owe m u c h to G E R A L D S U S S M A N , J O H A N DE K L E E R , G U Y S T E E L E , D R E W M c D E R M O T T a n d M A R V I N M l N S K Y for insp i ra t ion on the se top ic s . I a l so t h a n k J O S E P H S C H A T Z , P E T E R S Z O L O V I T S , R A N D A L L D A V I S , D O N A L D K O S Y , a n d M A R K F O X for v a l u a b l e d i scus s ions , a n d J A I M E C A R B O N E L L , J O H N M C D E R M O T T , a n d E D W A R D S H O R T L I F F E for s u e s t i n s u b s t a n t i v e i m p r o v e m e n t s to thi s p a p e r . T h i s research w a s s u p p o r t e d by t h e Defense A d v a n c e d R e s e a r c h P r o j e c t s A e n c y ( D O D ) , A R P A Order N o . 3597, m o n i t o r e d by t h e Air Force Av ion ic s L a b o r a t o r y under C o n t r a c t F 6 1 5 8 1 K 1 5 9 . T h e v iews a n d conclus ions conta ined in this d o c u m e n t are those of t h e a u t h o r , a n d shou ld n o t b e in te rpre ted as represent ing the official pol ic ies , e i ther e x p r e s s e d or impl ied , of the Defense A d v a n c e d R e s e a r c h P r o j e c t s A e n c y or the G o v e r n m e n t of the U n i t e d S t a t e s of A m e r i c a . The ''Probability Problem § 1 . R e c e n t succes se s of e x p e r t s y s t e m s s t e m from m u c h h a r d work of des igner s a n d exper t s in e l ict ing, encoding , a n d e x a m i n i n the n o r m a l l y t ac i t rules of r ea son ing employed by the e x p e r t s . T h e s e three t a s k s of e l ict ing, encod ing , a n d e x a m i n i n rules of rea son ing influence each other . T h e des igner m u s t b e ab le to encode the rules e l ic ted , t h e e n c o d i n m u s t fac i l i ta te e x a m i n a t i o n of the rules in o p e r a t i o n , a n d the s o r t s of in format ion el icted m u s t s e e m i m p o r t a n t to t h e e x p e r t l e s t . e x a m i n a t i o n s e e m point les s . M u c h current p r a c t i c e e m p l o y s so-ca l led j u d m e n t a l rules for e n c o d i n t h e i n f o r m a n t ' s rules of r e a s o n i n . J u d m e n t a l rules , for e x a m p l e t h o s e in M Y C I N , i n c o r p o r a t e bo th p r o p o s i t i o n a l a n d p r o b a b i l i s t i c in fo rmat ion . A n i n v e s t m e n t adv i so r s y s t e m , for in s t ance , m i h t b e iven a rule I F : 1. T h e cl ient ' s i n c o m e b r a c k e t is 5 0 % , a n d 2. T h e client careful ly s t u d i e s m a r k e t t r ends , T H E N : 3. T h e r e is ev idence (0 .8) t h a t the inve s tment shou ld be in h ightechnology . ( f rom [DAVIS 1979]) a n d would use the rule to d raw conclus ions whose cer ta in ty f a c t o r s d e p e n d on the o b s e r v e d cer ta inty f ac to r s of t h e h y p o t h e s e s (l ines 1 a n d 2) a n d the cer ta in ty factor (0 .8) of the rule itself. T h o u h few e x p e r t s y s t e m s a c t u a l l y t r e a t these numer ica l r a d e s of cer ta inty a s B a y e s i a n p r o b a b i l i t i e s a n d cond i t iona l p robab i l i t i e s , their in te rpre t a t ion usua l ly a p p r o x i m a t e s t h a t of B a y e s i a n probab i l i t i e s , n a m e l y a s s u b j e c t i v e degree s of belief. A l t h o u h s imple , these j u d m e n t a l ru les h a v e s u p p o r t e d d e v e l o p m e n t of m a n y i m p r e s s i v e a p p l i c a t i o n s . § 2 . In sp i te of t h e fruit fulness of th i s a p p r o a c h , its p r a c t i t i o n e r s expre s s d i s c o m f o r t wi th severa l of i t s r equ i rement s . One difficulty is t h a t while it is re la t ive ly e a s y to elict t e n t a t i v e p ropos i t i ona l ru les f rom e x p e r t s a n d from p e o p l e in genera l , it is cons iderab ly harder to ge t c o m m i t m e n t to p a r t i c u l a r r a d e s of cer ta inty . T h a t is, one ' s h u m a n i n f o r m a n t m i h t quickly s u e s t the p r o p o s i t i o n a l p a r t of t h e a b o v e rule IF 1 a n d 2, T H E N 3 b u t m i h t t ry to avoid a s s i gn ing 0.8 or any other n u m b e r to t h e rule . Worse still , ind iv idua l i n f o r m a n t s f requently vary in their answers to a r e p e a t e d quest ion d e p e n d i n on the d a y of the week, their e m o t i o n a l s t a t e , the p r e c e d i n ques t ions , and other e x t r a n e o u s f ac tor s . T h i s further a r a v a t e s t h e sens i t iv i ty of a n s w e r s to the p h r a s i n of ques t ions no ted by T V E R S K Y a n d K A H N E M A N . A n o t h e r difficulty s t e m s f rom the se . N o t i c i n the he s i t ancy of their i n f o r m a n t s , s y s t e m des igner s te s t the sens i t iv i ty of the s y s t e m ' s p e r f o r m a n c e t o the s e t of n u m b e r s u sed . R e p o r t e d e x p e r i m e n t s show the n u m b e r s to not ac tua l ly m e a n exac t ly w h a t they seem to m e a n , for the p e r f o r m a n c e of m o s t s y s t e m s r e m a i n s c o n s t a n t under al l s o r t s of s m a l l ( < 0 % ) p e r t u r b a t i o n s in the prec i se v a l u e s u s e d . U n d e r s t a n d a b l y , e x p e r t s y s t e m des igner s have difficulty ju s t i fy ing their use of the n u m e r i c a l j u d m e n t s in face of these ind ica t ions of p sycho log ica l a n d p r a m a t i c unrea l i ty . U n f o r t u n a t e l y , they h a v e h a d to s t i ck to their u n s , s ince no s a t i s f a c t o r y a l t e r n a t i v e h a s been a p p a r e n t . § . O n e hope for iv ing exper t s y s t e m des igners w h a t they w a n t is to i n v e s t i a t e their p rac t i ce , to seek s impler b u t e q u i p o t e n t s o r t s of in fo rmat ion and o p e r a t i o n s which c a p t u r e t h e g ra in of t r u t h in t h e p robab i l i s t i c a p p r o a c h . S o m e t h i n m i h t exp la in bo th the de s igner s ' in tu i t ions a n d their s y s t e m s ' succe s se s . S ince t h a t s o m e t h i n is c lear ly n o t s t a n d a r d B a y e s i a n p robab i l i ty theory or its ce r ta in ty fac tor v a r i a n t s , c an we find w h a t really under l ie s cur rent p rac t i ce ? T h i s p a p e r a n s w e r s Y e s . R e c e n t d e v e l o p m e n t s concerning r e a s o n e d a s s u m p t i o n s s u e s t an a p p r o a c h which a l lows fami l iar s o r t s of rules of r ea son ing , j u d m e n t s of cer ta in ty of conclus ions , a n d