Revision Log: 1 add preconditions
PROCESS: Pconditions: Revision Log:
heat-flow
not-heat-connection Tsource Tdestination Tpath heat-flow-aligned Tpath
1 negate preconditions
PROCESS : heat-flow
Pconditions: heat-connection Tsource Tdestination Tpath
not-heat-flow-aligned Tpath
Revision Log: 1 negate preconditions
PROCESS: heat-flow
Qconditions: greater-than (a (temperature Tsource))
(a (temperature Tdestination)) greater-than (a ((heat Tdestination))) (a (minimum-heat-point Tdestination))
Revision Log: 1 add quantityconditions
PROCESS : heat-flow
Qconditions: equal-to (a (temperature Tsource))
(a (temperature Tdestination))
Revision Log: 1 negate quantityconditions
PROCESS : heat-flow
Qconditions: less-than (a (temperature Tsource))
(a (temperature Tdestination))
Revision Log: 1 negate quantityconditions
PROCESS: Influences :
heat-flow
Revision Log: 1 delete influences
tween the different revisions is required. W ith increasing num bers of processes in the dom ain theory, the num ber of po tential revisions will increase proportionately. It is u n likely th a t em pirical evidence will always be available to discrim inate between com peting theories, and we m ust therefore look to other criteria for selecting one appropriately.
7.8
D iscu ssio n
7 .8 .1 R e la te d W o rk
M ost of th e systems introduced in C hapter 2 have only lim ited revision abilities. This is particularly tru e for BACON, STERN and HDD which concentrate on tren d detection in the discovery o f num erical laws. HDD does have a m echanism for attaching conditions to hypotheses in the event of failure, b ut this is very lim ited. There are a couple of systems, however, which do address revision in a sim ilar way.
R ajam oney’s work on COAST in pzirticular [91] is very closely related to the work here since it also uses Q ualitative Process Theory to represent its dom ain theory. He uses a scheme of abstraction in revision by which groups of proposed revisions are ab stracted a t a high level so th a t they m ay be subject to experim entation collectively. This allows th e refutatio n of m any revised theories before they are completely generated, and relies heavily on th e experim entation com ponent o f the system . A bstract revisions which cannot be differentiated on the basis of experim ents are then refined to concrete theo ries. The set of revision operators th a t is used by COAST, however, is not complete. It does not allow individuals to be added or deleted firom processes, and it does not allow individuals to be modified in any way either.
F urtherm ore, th e revision strategy adopted by COAST is very restricted and this also constrains the po tential revisions so th a t only a subset can be generated. COAST groups observations together and considers only revisions which can account for the whole group at once, ignoring revisions th a t involve m ultiple individual revision operator applications. By contrast, MED groups observations together in order to simplify the revision procedure, and to conserve resources. If no revision can be generated, MID progressively considers more com plicated revisions all the way through to those using observations individually. Thus we have considered a com plete set of revision operators, constrained only by w hat is syntactically and sem antically acceptable.
cation of background knowledge.
K arp ’s Hypgene program [46], which reasons in th e held of m olecular biology, also
uses a qualitative representation to model the dom ain theory, and provides a similar set of operators to those described here. He also m entions the possibility of including operators for modifying the class knowledge base o f the system (sim ilar to the background knowledge base here), b u t provides no details of how this m ight be im plem ented.
An zmalogous problem to theory revision is th a t of knowledge base refinement which involves the m odihcation of a knowledge base as opposed to a (scientihc) dom ain theory (see, for example, [28, 31]). Ginsberg [30] points out th a t m uch of this work on theory revision applies equally to knowledge base refinement. B oth involve th e m odihcation of a repository of knowledge in response to failures or inadequacies of some kind. We wiU not a tte m p t a detailed analysis of the similarities or differences here, however.
7 .8 .2 C o n c lu sio n s
Theory revision is a system atic process. W ithout th e use of selection criteria to constrain the space of revisions (which we consider in the next chapter), it is reduced to the syntactic m anipulation of dom ain theories. This in itself requires th a t a com plete set of revision operators exists, som ething which has been lacking in some previous systems. There are, however, constraints. Revised theories m ust be consistent w ith the observations and scenario th a t caused the theory failure to arise, and the revision operators m ust therefore be appropriately applied. Furtherm ore, the use of background knowledge in some form in order to facilitate the revision of theories, provides other requirem ents. The background knowledge itself should be capable of being revised, a t least in principle, for it is not inviolable, merely accepted w ith a greater certainty. Thus it should be m ore difficult to revise background knowledge, b u t not necesseirily im possible.
M H ) provides a facility th a t allows effective and efficient theory revision. It is sim ilar to the revision procedures in other systems th a t use sim ilar knowledge representations, bu t offers significant advances. By grouping observation, revisions are generated in a progressively m ore com plicated way, so th a t the sim plest revisions are considered first, and com plicated revisions later. While other systems consider only simple revisions, M H ) retains this advantage in term s of resources, b u t m akes the procedure complete by considering m ore com plicated revisions as and when necessary.