underlying abductive inference certainly is a first step toward mechanizing the invention of new or revised scientific laws and concepts. But these inferences cannot be demonstrated to be the inference which the scientist followed to the new notion, rather, only an inference having this new notion as its conclusion. Thus, it is not at all clear that a theory of abduction adequate for machine intelligence applications must await a full understanding of human cognition. Quite the contrary; the preferred approach is to attempt to develop a theory of abductive inference on the basis of a direct logical analysis, retreating to the more fundamental problem of human cognition only if the techniques of logical analysis fail.
To consider what might be expected from a direct attack on the logic of abduction a brief characterization of inferential steps constituting such inference is presented below. This characterization takes the viewpoint of some unspecified knower "X," either a scientist or an abductive machine intelligence system.
(1) X is surprised while using theoretical structure T by some occurrence, 0, because 0 is not among X's set of expectations that are based on T.
(2) X represents 0 by a determinate set of data, D.
(3) X demonstrates that D is more than simply unexpected; it is anomalous in the sense that T predicts not-Z).
(4) X traces not-Z) back to those components [?] ,7*2,...] of its total theoretical structure which entered directly into Ts prediction of not-?>.
ABDUCTIVE PROCESSOR = ( AB^UCT.V^NFERENCES Figure 3.9.- Abductive inference.
(5) X determines which element, Tj, in [7^ ,?2,...] is the most likely "villain" behind X's misexpectation.
(6) X attempts to reformulate Tj in such a way that when the new Tj* is substituted for Tj in a revised T*, 0 can be represented by D* which, in turn, is predicted by T*. (If successful, the next step is (9) below.)
(7) If not successful, X repeats steps (5) and (6) above with the remaining elements of [Tx ,T2 -???] in order of decreasing likelihood until all possibilities are exhausted. (If successful, the next step is (9) below.)
(8) If still not successful, X repeats steps (5) and (6) with the remaining elements of T, in order of increasing theoretical content and scope, the last component tried being the fundamental "deep structure" model itself.
(9) X makes all adjustments in T* necessitated by the adoption of Tj*, including generating a new set of expectations 0*.
(10) X uses T* until the next "surprising" occurrence.
This characterization of abduction, though not as detailed and precise as that which would result from further investigation, is precise enough to suggest three key problems standing in the way of mechanized abductive inference. First, how should 0 best be represented as data so that later re-representation is facilitated, and how should these re-representations be performed? Second, is the initial selection of "villains" best achieved by parallel search, hierarchical serial search, or some other technique? Third, can the formulation of the Tj* replacement of Tj be captured in a stepwise inference in which preceding steps uniquely constrain the selection of the next succeeding step, or must some other technique be used? Note that all of these may be addressed on logical grounds, independent of the broader questions of human cognition.