--- title: "On the Mechanization of Abductive Logic" authors: - "Harry E. Pople" doi: "" url: "https://ijcai.org/proceedings/1973/20" pdf: "https://ijcai.org/proceedings/1973/0020.pdf" venue: "Proceedings of the Third International Joint Conference on Artificial Intelligence" year: 1973 pages: "147-151" track: "Theorem Proving and Logic: 2" keywords: --- # On the Mechanization of Abductive Logic Harry E. Pople *Proceedings of the Third International Joint Conference on Artificial Intelligence*, pages 147-151. ## Abstract Abduction is a basic form of logical inference, which is said to engender the use of plans, perceptual models, intuitions, and analogical reasoningall aspects of .intelligent behavior that have so far failed to find representation in existing formal deductive systems. This paper explores the abductive reasoning process and develops a model for i ·ts mechanization, which consists of an embedding of deductive logic in an iterative hypothesis and test procedure. An application of the method to the problem of medical diagnosis is discussed. ## Full text ## Introduction There has been growing criticism lately concer­ ning the methodology of artificial intelligence. While differing in the specifics of their analyses of the problem, most thoughtful observers seem to feel that the current stock of deductive machinery is simply not up . to the task at hand. A deficiency often cited with regard to the present deductive procedures is their inability to represent and make use of the sort of plans, perceptual models, intu­ itions, and analogical reasoning processes that characterize at least some phases of virtually all human problem solving activity.l,2,3 As remedy, various authors have suggested replacement of the existing formalisms .by alternative representations such as higher logics, analog models, or general purpose programming languages . In this paper, we offer yet another diagnosis of the problem, and propose a somewhat different solution. In our view, -the principal deficiency of existing systems is their reliance on a single form of logical inference - deduction - which, though essential, is inadequate for many types of problem solving activity. Our proposed remedy is to extend the existing formal systems to embrace additional forms of inferential reasoning, especially that of abduction. Abductive inference is one of the three funda­ mental modes of logical reasoning - the others being deduction and induction - characterized by Peirce4 as the basis of scientific inquiry. According to Peirce, abduction underlies "all the operations by which theories and conceptions are engendered;" moreover, "abductive inference shades into percep­ tual judgment without any sharp demarcation line between them." By his account, it is "the only kind of reasoning that supplies new ideas, the only kind which is, in this sense, synthetic." Finally, ac­ cording to Peirce, abduction plays a key role in the process of an\_alogical reasoning . Since these are the very aspects of intelligent behavior that have been found wanting in the method­ ology of artificial intelligence, it has seemed to Harry E. Pople, Jr. Graduate School of Business University of Pittsburgh Pittsburgh, Pennsylvania 15260 us important to explore the abductive reasoning process and to .develop a model for its mechanization. The essence of abductive inference is the generation .of hypotheses, which, i f true, would ex­ plain some collection of observed f acts. This con­ cept is illustrated by the following example from McCullochS: " . .. abduction starts from the rule and guesses that the f act is a case under that rule: All people with tuberculosis have bumps; Mr. Jones has bumps; perhaps Mr. Jones has tuberculosis. This, sometimes mistakenly called 'inverse probability ,' is never certain but is, in medicine, called a diagnosis or, when many rules are con­ sidered .a differential diagnosis, but it is usually fixed, not. by a stat i stic, but by finding some. other observable sign to clinch the answer." It is important to observe that the hypothesis formu­ lated in an abduction is t ypically "quite different from anything observed, something that is in the majority of cases empirically unobservable." Thus the abductive generation of hypotheses is clearly set apart from that of induction, wherein we t ypically "generalize from a number of cases of which something is true, and infer that the same thing is probably true of a whole class."6 The various modes of logical inference can be characterized as alternative forms of argument based on the following syllogistic schema: I. Major Premise (rule) V (x) (P(x) ~ Q(x)) II. Minor Premise (case) III. Conclusion (fact) P(a) Q(a) ## Specifically: - A) The process of deduction reasons from a major premise (I) and a minor premise (II) to a conclusion (III); thus, from the rule "all things that are P are also Q" and the specific .case "~is a P", we conclude via deduction inference that "~is a Q." - B) Induction, on the other hand, comprises the process of reasoning from a case (II) and a fact relating to that case (III) to a hy­ pothesis .of the rule (I) that describes this relation in general terms. Thus from an observation "a is a P" and the fact "a is a Q," one for;;;- of inductive inference­ would be to hypothesize that "perhaps all things P are also Q." - C) The process of abduction is characterized by the form of reasoning that takes a given fact (III) in conjunction with a rule (1) to hypothesize a case (II) that could account for or "explain" the observ.ed fact. Thus, from an observation "a is a Q" and the rule "all things Pare also Q", one might hy­ pothesize, via abductive inference, that "perhaps~ is a P." These forms of reasoning all entail the same underlying representation of information. Whatever language is used to express the axioms and conjec­ tures of a deductive system can also be used to ex­ press the observations and hypotheses required for induction or abduction. However, although they em­ ploy a common information structure, each of these inference schemes requires the development of a separate computational procedure for its mechaniza­ tion. Some work has been reported in the literature dealing with such procedures; these have, for the most part dea7t with approaches to the mechanization of induction, · 8 although there has also been report of at least one abductive system that takes a 'British Museum' approach to the generation of all possible hypotheses.9 The real problem in this area, of course, is the selective generation of hypotheses that have some reasonable prospect of being valid. Our approach to this problem has been guided by our interpretation of the traditional maxim of Occam's Razor, which requires adoption of that hypothesis which is the simplest - in the sense that it contains the smallest number of independent types of elements, adding least to what has been observed. We have taken this to mean that the preferred hypothesis is the one that reflects to the greatest extent possible a synthesis of ideas evoked from the observations. This heuristic criteri­ on of acceptability will be discussed f urther in a later section, where procedures for testing and vali­ dation of hypotheses will also be described. Before pursuing these matters in more det ail, it will be necessary to develop first the framework within which hypothesis formation takes place in our system. It will be shown in the following section that the process of abductive inference can be ac­ complished by means of a procedure that makes use of much of the machinery already developed for deductive logic. Although our abduction procedure has been implemented using GOL10as the basic deductive pro­ cessor, for purposes of explication we describe the procedure in the context of the better known deduction scheme of S-linear resolution.ll In the following, we review first the linear de­ duction method, recasting the exposition somewhat so that the intrinsic problem-reduction structure of the method is exposed . In this context, we present the notion of partial trees as a way of characterizing the AND/OR search procedure underlying this form of deduction. We then show that abductive hypothesis formation can be effected by a synthesis process that merges partial trees into complete abduction graphs. ## 1. Linear Resolution Viewed as a Problem Reduction Procedure In this section, we provide a somewhat unorthodox description of the axioms, conjectures, and inter­ mediate structures that arise in the course of a typical linear deduction. Our purpose is to show that linear resolution can usefully bl viewed as a problem­ reduction procedure (cf . Nilsson 2 ), which uses back­ ward search to develop direct derivations of conjunc­ tive sets of subgoals . It is a relatively small step then to show how the same sort of search, operating on the same form of data, can be used abductively to develop direct 'explanations' of conjunctive sets of observations. ## 1.1 The Information Structure A deduction problem is typically stated in terms of a collection of axioms A and a conjecture C which is to be deduced from the axioms. It is con­ ventional in resolution based systems to conjoin the negation of C to the set A, convert the resulting set of expressions to quantifier free form , then to search for evidence of a contradiction that would render the expression AAC unsatisfiable. An alternative formu­ lation of the problem would be to search for direct proof of the disjunction AVC. In this section, we discuss the procedure from yet a third point of view - namely, the demonstration of AI- C; i.e ., taking the expressions of A as axioms, to derive C by means of a direct proof procedure. These are all essentially equivalent views of the problem; they merely require somewhat different handling of the conversion to quantifier free form and subsequent interpretation of the information structure . The first step in our procedure is to convert the axiom set A to conjunctive-normal quantifier-free form, using the techniques first outlined by Davis .l3 Or­ dinarily, the conjecture C would then be negated and added to this axiom set. In our case however, as in the direct proof procedures of Cooperi4 and Maslovl5 the conjecture is transformed into disjunctive normal form on the basis of the following revised rules: 1. elimination of implication; (as described by Davis) 2. reduction of the scope of negation; (as in Davis) 3. replacement of each universally quantified variable by a skolem function having as arguments the vari­ ables of any existential quantifiers occurring before the universal quantifier. 4. 4 . elimination of existential quantifiers. 5. transformation of the resulting expression to disjunctive normal form . The result of this process will be a disjunctive ex­ pression consisting of one or more conjunctive sets of literals. Consider for example the following set of axioms: 1. 'v'x(P(x)A Q(x) :::> R(x)) 2. Vw(P(w).::>R(w)) 3. \iyP(y) and the conjecture: Converting the axioms, we obtain the clauses: 1. {P(x), Q(x), R(x)} 2. {P(w), R(w)} 3. 3 . {P(y)} and the conjecture is transformed into the following disjunctive form: where 'a' is a constant skolem function. Each disjunct in a conjecture can be considered to be an alternative formulation of the derivation problem, as illustrated by the following: Assume that the converted conjecture comprises n disjuncts (Dl, D2, ··· Dn). For each Di, we can formulate the expression: (Bi V Di), where Bi repre­ sents the disjunction of all Dj for j f i. This ex­ p~ession ca~ then be converted to implicative form (Bi J Di); Bi can be added to the axiom set A; and the problem 1s then reduced to derivation of the con­ junctive expression Di. One such alternative state­ ment of the problem can be formulated on the basis of each Dk in the conjecture. In the example above, the conjecture would give rise to the two alterna­ tives: Prove: {Q(a)} II. Assuming: 1. {P (x), Q(x), R(x)} 2. {P (w), R(w)} 3. {P(y)} 4'. {Q(a)} Prove: {R(z), Q(z)} These formulations of the problem are surrogates in the sense that the solution of either one of them is equivalent to solution of the original problem. Collectively, therefore, they constitute an "OR" branch in the search tree, each successor of which is an "AND" node (consisting of the conjunction of liter­ als in the conjecture associated with that sub-prob­ lem). Any unit clause in the axiom set is considered to be unconditional assertion of fact; all others are conditional. For example, the clause {1 1 , 1 2 , ·· ·-In} can be thought of as asserting the implication (M :::> 1i), where M is the conjunction of all Lj for j f i. This can be interpreted as expressing a con­ dition (namely M) under which 1i can be deduced (by means of Modus Ponens perhaps with substitution). Similar implicative expressions can be developed to assert the conditions under which each of the other literals of the clause can be derived. These implicative expressions can be thought of as productions, or 'rewrite rules: that are the transformations by which a problem may be 'reduced' - i.e., replaced by an equivalent set of sub-problems. In the running example, if we reformulate the axioms to emphasize the conditional nature of axioms 1 and 2, the resulting set of productions would be: 3. {P(y)} here the symbol '+' (actually 'c:') is to be thought of as a replacement operator. ## 1.2 Interpretation of the deduction procedure The sub-problems described in the preceding section correspond in an obvious manner to the various ways in which an initial clause may be selected from the set of support (where this consists of clauses arising from the conjecture) in a resolution deduc­ tion. It should .also be readily apparent from a com­ parison of the two procedures that the problem-re­ duction search process described below, which results from our interpretation of the axioms as productions, is essentially the same ag the S-linear resolution method described by 1ovelandll. A linear deduction is a procedure that 'solves' the conjunctive. problem posed by a conjecture by systematically dealing with each of the conjuncts (sub-problems) in turn. Proceeding typically from right to left, each literal of the conjecture is processed in the following \_manner: First, the literal is 'matched' (using Robinson's unification procedure 16) against each unit clause and against the literal to the left of '+' in each pro­ duction of the axiom set. Every successful match gives rise to an alternative successor of the starting conjecture. Any match that succeeds with respect to a unit (unconditional) axiom provides a solution to the sub­ problem posed by the subject literal (which is true 'by assumption'). Thus this term may be deleted from the list of literals of the successor conjecture (af­ ter appropriate instantiation of any bindings of variables established in the matching process). For those matches that succeed with respect to the left hand side of a production (conditional axiom), the successor conjecture is formed by replacing the subject literal by the appropriately instantiated collection of literals comprising the right hand side of that axiom. The set of successors constructed as above, taken collectively, constitute an 'OR' branch in the search tree - each element of which is a surrogate for the original conjecture. If a successor node is null (i.e., contains no further sub~problems), the portion of the search tree by which it was discovered constitutes a solution to the original problem. A non-null node with no successors cannot be solved. It sometimes happens that while developing the search tree for a particular literal (sayS), a sub­ goal is generated that is the negation of that which is to be proved (in this case:~). In such cases, it can be shown that, provided that all other conjunctive subgoals in the tree are satisfied, a proof of S has been established by reductio ad absurdum, and that the contradictory subgoal can be deleted from further con­ sideratio~ immediately upon its generation. This refinement to the basic problem-reduction scheme can be seen to be related to the subsumption condition which is required for completeness of linear resolution strategies. To illustrate the linear deduction proced~re, we give below a derivation of version II of the previous example. The starting form of the conjecture is: {R(z), Q(z)} To show that this is derivable we must now show that each literal is either an instance of an axiom (in which case it can be deleted from the set of sub­ problems) or else is implied by some conjunction of literals, each of which is derivable. In order to prove 'Q (z)' we can use Axiom 4 '·. By substituting 'a' for the existentially quantified variable 'z' of the conjecture, we see that the right-most sub-problem in the set {R(a), Q(a)} is solved trivially, and we are left with the successor problem {R(a)}. Since the literal 'R(a)' matches two of the con­ ditional axioms, there will be two successor nodes: a: {P(a), Q(a)}, and b: {P(a)} Proceeding with 'a,' we find again that the right most literal is an instance of a unit axiom (4') and is therefore solved trivially. The successor {P(a)} is then solved by matching with axiom 3. Since the successor is null, this completes the deduction. The situation described previously, where a sub-goal is generated that negates one of its ances­ tors- thereby giving rise to a 'reductio' proof - is illustrated by the further elaboration of alternative 'b' above, which is left to the reader. ## 2. A procedure for the Mechanization of Abduction In a deduction, the objective is to determine whether a given conjunctive set of statements (repre­ senting conjectures) is true. In an abduction, on the other hand, the objective is to determine ~ the conjunction (now representing observations) is true. There are substantial similarities in the processes involved in answering these two types of questions; however, there are also significant differences. In answering the 'why' question, it is obviously important to be able to determine 'whether;' thus de­ duction may be considered to be a process subordinate to abduction. If deduction succeeds in 'deriving' a conjunctive statement of observations on the basis of a given set of axioms, the 'why' of abduction is answered by the resulting solution tree. Note that alternative derivations, which by and large are con­ sidered to be irrelevant in answering the deduction question, may be important in abduction - since they constitute alternative explanations of the data. There is another important respect in which . ab­ duction differs from deduction; whereas deduction accepts 'no' for an answer, this is not considered an acceptable reply to the 'why' of abduction. Thus it is typically necessary for the abduction process to generate additional hypotheses, which if added .to the axiom set, would allow deduction to respond 'yes.' Again, it is obvious that there will be, in general, many ways in which such 'explanations' of the data may be generated. As candidate hypotheses, our procedure selects those literals that are abandoned by deduction, in the sense that they fail to generate successor nodes. A candidate hypothesis is entertained seriously if it arises in the partial search trees that are developed on the basis of two or more of the data .making up the conjunctive observation being explained; the more data accounted for by a candidate hypothesis, the more highly regarded it is by the abduction processor. This is our implementation of the maxim of Occam's Razor. Data contribute to a hypothesis by means of an operation that we refer to as synthesis, which is actually analogous to the operation of factoring in resolution. This process of synthesis, or fac­ toring across partial trees, is at present the only mechanism by which abductive hypotheses are generated in our system. Other useful heuristic criteria will undoubtedly be forthcoming, but what form these may take is still an open question. The combined abduction/deduction procedure is illustrated by the following example, based on our application of the method to the problem of clinical diagnosis. Assume that we have available a pathophysiologic­ al data. base, structured along the lines suggested in the preceding section, that includes the following sort of axioms: (note that the replacement operator "+" should rightfully be interpreted as "could be caused by" rather than as "implied by" in this context.) ## Axioms Al: {chills + presence (P, S) 1\ inflammatory (P)} A2: {pain (R) +presence (P,S) A Located-in (S,R)} A3 : {inflammatory (abscess)} A4: {located-in (liver, right-upper-quadrant)} AS: {jaundice+ presence (P, liver)} Here, we have indicated only those conditional axioms that are considered relevant for the application at hand. Other variations of the conditional axioms Al and A2 which are not useful for purposes of diag­ nosis are not explicitly displayed. (In our work, this form of information structure is implemented by means of the GOL DELTA function (cf, Isnerl7.) Assume that the following conjunctive set of symptoms has been observed: {chills, pain (right-upper-quadrant)} The diagnostic task is to formulate some hypothesis of the form: {presence (P,S) } where P is some abnormal process, and S is some struc­ ture such that the presence of P at S could account for the observations given. Using the synthesis criterion suggested above to control the adduction of hypotheses (which are dis­ played to the left of the vertical bar in lines 4-6 below) a diagnostic model of the pathology can be derived as follows: | 1. | {jchills, pain (right-upper­ quadrant} | observation | |------|---------------------------------------------------------------------------------------------|---------------| | 2. | { jpresence (X,Y), located- in (Y, right-upper quadrant), chills} | 1, A2 | | 3. | { jpresence (U,V), inflammatory (U), presence (X,Y), located­ in (Y, right-upper-q~adrant)} | 2, Al | | 4. | {presence (U,V) I inflammatory (U), located-in (V, right-up­ per quadrant)} | 3, synthesis | | 5. | {presence (U, liver)j inflam­ matory (U)} | 4, A4 | | 6. | {presence (abscess, liver)j} | 5, A3 | Except for the introduction of a tentative hypothesis in line 4 above, and the fact that literals are processed on a first-in-first-out basis, this derivation follows the usual form of a linear deduc­ tion as previously described. The abduction task is completed when no remaining subgoals occur to the right of the vertical bar; any terms occurring to the left of the bar at that juncture constitute the basis for a diagnostic model. It may be of some interest to follow the course of development of the diagnosis in the case illus­ trated above. In line 4, where the synthesis step occurs, what is hypothesized is that some as yet unknown process affecting an unspecified structure is responsible for both of the observed symptoms. The instantiations that then take place in the final two steps of this derivation entail contributions from the search trees developed from each of the root nodes; thus the proposed diagnosis represents a true synthetic inference. A hypothesis developed on the basis of such a procedure is, in general, not a unique explanation of the data, and the problem then becomes one of discrimination, among contending diagnoses. This phase of the problem entails the use of deductive, as well as abductive logic. Once a diagnostic model has been proposed, it can be used to generate pre­ dictions of additional unreported manifestations of the assumed pathology. Thus, for example, in the case illustrated above, the presumption of a liver abscess can be used to deduce the prediction of jaundice (on the basis of axiom 5). Such predictions can then be subjected to empirical verification. Any new observational data can then be fed back via the abduction procedure through another iteration of the cycle - giving rise, perhaps, to revised hy­ potheses that may in turn generate new predictions, leading to new observations, and so forth. ## 3. Discussion The iterative hypothesis and test procedure out­ lined in the preceding section is, of course, one of the basic paradigms of human cognitive and problem solving activity. The performance of any task that is basically synthetic in nature entails the use of this procedure. We would include in the list of such tasks those episodes of comprehension and planning that arise in the course of any real problem-solving act. A number of studies are presently being con­ ducted at Pittsburgh to evaluate and further develop this concept. These include the biomedical theory formation project which has been reported previously, and a study of planning in Qrganic synthesis, re­ cently completed by Smith, 1 ~which uses abduction to develop strategies for deductive planning of a syn­ thesis. Another interesting project, now nearing completion, uses abductive logic to develop per­ ceptual models for use in natural language compre­ hension; another exploits the planning and model­ building capabilities of the system in the automatic programming task environment. There would appear to be a number of ways to go from here. As practical systems of higher logic become available, we may want to raise the sights of the abduction processor accordingly; our GOL imple­ mentation already has limited 'higher level' capa­ bilities that have been found extremely useful in some applications. Certainly, the procedure should be extended to include a mechanization of induction as well as the other two forms of inference: reasoning by analogy, according to Peirce, consists of an induction and an abduction followed by a deduction. Additional applications that may shed further light on the processes involved, such as procedures for strategic planning in theorem proving, should be investigated. While we have in no way begun to exhaust the questions and issues that this new methodology raises, results from our initial forays into the field have encouraged us to continue with the task at hand. ## References 1. Sloman, A., Interactions between philosophy and artificial intelligence: the role of intuition and non-logical reasoning in intelligence. Ar- tificial Intelligence (2), (1971), pp. 209-225. 2. Anderson, B. and Hayes, P., An Arraignment of Theorem Proving or The Logicians' Folly, Dept. of Computational Logic, University of Edinburgh. DCL Memo #SO, 1972. 3. Dreyfus, H.L., What Computers Can't Do. Harper & Row, 1972. 4. Peirce, C.S., Collected Papers of Charles Sanders Peirce. C. Hartshorne, P. Weiss, and A. Burks (eds.) 8 vols., Cambridge, Mass., 1931-1958, es- pecially Vol. II, p. 272-607. 5. McCulloch, W.S., "What's in the brain that ink may character1" Presented at the International Congress for Logic, Methodology, and Philosophy of Science, 1964. Reprinted in: W.S. McCulloch, Embodiments of Mind, The MIT Press, 1962. 6. Goudge, T.A., The Thought of C.S. Peirce, Dover, 1969. 7. Plotkin, G .D., A further note on inductive gen- eralization. In: Mach. Int. 6, Meltzer, D. and . Michie, B. (eds.), Amer. Elsevier, 1971. 8. Popplestone. R.J., An experiment in automatic in- duction. In: Mach. Int. 5, Meltzer, B. and Michie, D. (eds.), Amer. Elsevier, 1970. 9. •. Morgan, C.G., Hypothesis Generation by Machine. Artificial Intelligence 2 (1971) pp. 179-187. 10. Pople, H. E. , Jr., A goal oriented language for the computer, in Representation and meaning - Experiments with information processing systems. 11. Loveland, D. W. , A uni fying view of some linear Herbrand Procedures. JACM , Vol . 19, 2 , April , 1972 . 12. Nilsson, N.J., Problem Solving Me thods in ar- tificial intelligence . McGraw-Hill, New York, New York 1971. 13. Davis, M., Eliminating the irrelevant from mechanical proofs, Proc. Symposia in Applied Maths, Vol . XV, AMS. (Ed. Metropolis, et . al.) 1963, pp. 15-30 . 14. Cooper, D. C., Theorem Proving in Computers, in Advances in Programming and Non-Numerical Computation, L. Fox (ed.), Pergamon Press, 1966. pp. 155-182. 15. Maslov, S.Ju., Proof-search strategies for meth- ods of the resolution type. In : Mach . Int . 6, Meltzer , B. and Michie, D. (eds.) Amer. Elsevie~ 1971. 16. Robinson, J.A . , A machine-oriented logic based on the resolution principle . J. Ass. Comput. Mach. , 12, pp . 23-41 . 17. Isner, D.W., An inferential processor for inter- acting with biomedical data using restricted natural language. Proc. SJCC, 1972. 18. Pople, H. and Werner, G., An information processing approach to theor y formation in bio- medical research. Proc. SJCC , 1972 . 19. Smith , G., Jr. Mechanized procedures for strate- gic planning in synthetic organic chemistry, Doctoral Dissertation, Dept . of Chemistry, Univ . of Pittsburgh, 1973.