Putting Theories Together to Make Specifications

Representing knowledge w i t h i n an AI program AI programs are commonly conceived to embody knowledge, whether as program or as data . Procedu r a l embedding of knowledge may promote e f f i c i e n c y , and it may enable one to use e x i s t i n g program s t r u c t u r i n g techniques to impose some order on the embedded knowledge. However procedural embedding has disadvantages o f i n f l e x i b i l i t y , and i t makes i t d i f f i c u l t to incorpora te new knowledge, whether inpu t or from induc t i ve l e a r n i n g . Much of the d i squ ie t w i t h knowledge held as data seems to us to stem from i t s lack of s t r u c t u r e , a large c o l l e c t i o n of axioms or f a c t s unorganised, s lowing processing down w i t h i r r e l e v a n t i n fo rma t i on . Again we are exhorted to consider our commonsense knowledge of the wor ld as composed of a la rge number of m ic ro theor ies about p a r t i c u l a r * Mosses (1977) makes a s t a r t in t h i s d i r e c t i o n . aspects. But we are l e f t l a r g e l y in the dark as to how to put these mic ro theor ies together . This quest ion o f ' p u t t i n g theor ies together ' i s c lose to the hear t of our concern. Thus al though we cannot yet speak from experience, we very much hope tha t our theory b u i l d i n g techniques may eventua l l y g ive some f resh i n s i g h t i n t o the appropr iate organ isa t ion of knowledge in AI programs. A way of present ing theor ies so tha t people can understand them might help us to see how machines can make use of them.