REPRESENTATION AND USE OF EXPLICIT JUSTIFICATIONS FOR
KNOWLEDGE BASE REFINEMENT
Reid G. Smith Howard A. Winston
Tom M. Mitchell1Bruce G. Buchanan2
Schlumberger-Doll Rearch
Old Quarry Road
Ridgefield, CT 06877-4108
USA
From:Proceedings of the Ninth International Joint Conference on Artificial Intelligence, pp. 673-680, Los Angeles, CA, August 1985.
ABSTRACT
We discuss the reprentation and u of justification structures as an aid to knowledge ba refinement. We show how justifica-tions can be ud by a system to generate explanations – for its own u – of potential caus of obrved failures. We discuss specific information that is ufully included in the justifications to allow the system to isolate potential faulty supporting beliefs for its rules and to effect repairs.
This rearch is part of a larger effort to develop a Learning Ap-prentice System(LAS) that partially automates initial construction of a knowledge ba from first-principle domain knowledge as well as knowledge ba refinement during routine u. A simple im-plementation has been constructed that demonstrates the feasibility of building such a system.
I. INTRODUCTION
A Learning Apprentice System[6] is an interactive aid for building and refining a knowledge ba. Its aims are twofold: (i) to partially automate initial construction of a knowledge ba by generating shallow rules automatically from an approximate domain theory: and, (ii) to interact with urs to help refine the knowledge through experience gained during normal problem solving. In this paper, we concentrate on knowledge ba refinement. In our scenario, for problems where the Performanc
e Program (i.e., the com- ponent of the knowledge-bad system that performs the problem solving) fails to make an important inference, or makes an incor- rect inference, the ur advis the LAS of the failure. The system tries to explain why the failure has occurred (perhaps via a focusd interaction with the ur), and repairs or extends its knowledge ba as required.
This type of "knowledge acquisition in context”has previously been ud to advantage in TEIRESIAS [3, 1], in the context of the MYCIN system. Its advantages include: (i) it happens in the normal cour of u of the system, driven by failures, and therefore has the potential of tapping the knowledge of specialists without plac- ing large additional demands on their time: and, (ii) becau the specialist and system are working step by step through a specific problem, the learning is focusd on a relatively small portion of the knowledge ba.
Our main focus here is explanation of failures – assignment of blame to specific items in the knowledge ba. In terms of the ___________________
1
T. Mitchell is with the Computer Science Department. Rutgers University, New Brunswick. NJ 08903.
2
B. Buchanan is with the Computer Science Department, Stanford University Stanford, CA 94305, general model of learning systems prented in [2], this is the task of the Critic. Like TEIRESIAS, we u the reasoning trace exhibited by the Performance Program to help focus the interaction be- tween the LAS and the ur. However, instead of relying on the ur to explain why a particular rule or t of rules might have failed, the LAS us a type of dependency network – a justification structure – to construct possible explanations for the obrved failure.
The reasoning trace provides support for conclusions drawn by the Performance Program. The content of that support is the chain of rule firings that led to the conclusions. A justification structure, on the other hand, goes one step further. It records the support for an individual rule. The content of the justification is the deeper domain knowledge from which the rule is derived, together with the assumptions and approximations that have been ud in the derivation. For the purpo of refinement, the primary information captured in a justification structure is the manner in which errors propagate across dependency links. By using this information in concert with a taxonomy of error types, it is possible for the LAS to determine a constrained t of suspect supporting beliefs that can explain an obrved failure of some supported belief. This infor-mation is uful for focussing a dialogue between the LAS and the specialist. It is also esntial input to an automated revision sys- tem.
Several rearchers [5. 8] have studied methods for recording logical dependencies between beliefs, and for using such depen-dencies to guide inference. The current work extends the con- cepts of truth maintenance and dependency-directed backtrack- ing in that domain-specific knowledge about how to propagate errors through a dependency network is ud in addition to knowledge of the logical structure of the network itlf.
The XPLAIN system of Swartout [10] also ud justifications to generate explanations of its behavior for human urs. By con-strast, we focus here on using justifications to generate explana-tions for the system itlf of potential caus of obrved failures. This has led us to concentrate on determining precily which information is ufully included in the justifications to allow the system to isolate potential faulty supporting beliefs for its rules and to effect repairs.
We are using the Dipmeter Advisor*system in order to test our ideas in the context of a task (and Performance Program) that is already well understood [9]. We have implemented subprograms to extend this system and have undertaken some analysis of the generality of the method.
___________________
*
Mark of Schlumberger
A. GEOLOGICAL INTERPRETATION: THE
DIPMETER ADVISOR SYSTEM
The general task of the Dipmeter Advisor system is to infer the geological structures penetrated by a single borehole. For ex-ample, it might conclude that a late fault with strike 92°, has been penetrated at a depth of 1050 feet. Its primary input data is the dip or tilt of rock formations penetrated by the borehole, indexed by depth. It also us data from a variety of other logging tools as well as information about the local geology.
The system divides the interpretation task into an ordered ries of subtasks involving pattern detection, data aggregation, and abstraction. After each subtask has been completed, the ur is given the opportunity to examine, delete, or modify conclusions reached by the system. He can also add his own conclusions. In addition, he can revert to earlier stages of the analysis and repeat various subtask chains with different assumptions or data.
The rules in the system are empirical associations applied by the rule interpreter in distinct rule ts
羊剪绒大衣using a forward-chaining con- trol strategy. A simplified version of a rule ud to determine that a previously detected normal fault is a late fault is shown in Figure I-1.
Rule: NFR12
IF
there exists a normal fault pattern (p), and
there exists a red pattern (p1).
such that the length of p1 < 50 ft., and
such that p1 is above the fault plane pattern of p.
THEN
specialize p to be a late fault pattern
Figure I-1: Late Fault Rule
The geometry and dip patterns upon which NFR12 is bad are shown in Figure 1-2. A rough justification for the rule is as follows: A late fault typically has a distortion region directly above the fault plane. The distortion region for a late fault is generally thin be- cau the surrounding rock has been compacted, and is therefore not very plastic, at the time of faulting. The thin distortion region, projected onto the borehole, is manifested in dipmeter data as a short red pattern. The 50 ft. threshold is dependent upon as- sumed values for the distortion, region thickness and the fault angle. The rule assumes that dipmeter patterns can be detected in the vicinity of the fault. This will be true if bedding is prerved, the borehole is not washed out (caved in), and if the tool is operat-ing correctly. If, for example, the borehole were washed out over part of the distortion region, the length of the measured red pat- tern would be less than expected.
We u this rule as an example throughout the paper. We show how its detailed justification can be ud to isolate faulty support- ing beliefs that can prevent it from firing in some situations.
II. DESCRIBING ERROR TYPES AND RULE JUSTIFICATIONS The Critic us the type of error that the Performance Program has made, along with its reasoning trace, the current rule ba, and justification structures to guide generation of plausible ex-planations of a failure and to focus on plausible repairs.
Figure I-2: Late Fault Geometry
A. ERROR TYPES
In examining the justification of an incorrect rule, it is uful for the Critic to distinguish among vera
l class of errors becau: (i) it provides an additional constraint on the arch for suspected supporting beliefs; and, (ii)different class of errors suggest dif-ferent repair strategies. The error class currently ud by our prototype LAS for beliefs are shown in Figure II-1, and break down
into three main class:
Figure II-1: Types of Belief Errors
Rule Belief Error: Beliefs that are implications may suffer from one of the following class of error: OverGeneraiLeftHandSide (OGLHS) – the rule applies in situations in which it should not: OverSpecificLeftHandSide (OGLHS) – the rule does not apply in situations in which it should: OverGeneralRightHandSide (OGRHS) – the rule fails to make asrtions when it should or makes asrtions that are too general (e.g., drawing a conclu- sion about a normal fault where a conclusion about a late fault, a specialization of normal fault. is warranted); and, OverSpecificRightHandSide (O
SRHS) – the rule either makes asrtions when it should not or makes asrtions that are too specific (e.g., drawing a conclusion about a late fault where only a conclusion about a normal fault is warranted).梨子的种类
Numeric Parameter Error: A belief about the value of a numeric parameter may be incorrect by being an OverEstimate (OE) or UnderEstimate (UE) of the correct value of the parameter. Symbolic Parameter Error (SPE): This is an error regarding beliefs which are neither implications nor asrtions about numerical parameters. They simply have a truth value. For ex-ample, an erroneous asrtion that a comparator is "<" when it should be ">" is a symbolic parameter error.
We believe this error taxonomy to be uful in a variety of domains. It is also incomplete. We have not yet formalized, for example, errors in symbolic parameters for which a partial order can be established. In such cas, OverSpecific and UnderSpecific may be appropriately added as subclass of SymbolicParameterError.
B. JUSTIFICATION STRUCTURES
We define a justification structure to be a network of Beliefs con-nected by Justification Links. Beliefs reprent asrtions and Justification Links record the derivational dependencies between beliefs.
A Belief contains Asrtion, Type Of Belief, and Degree Of Belief slots (in addition to its links in the justification structure). The Asrtion is the actual statement of the belief. The Type Of Belief is one of definitional (i.e., no further justification is required), theoretical (i.e., bad on the half-order theory – could be incor- rect if the theory is incorrect), statistical (i.e., justified by statistical experience), or default (i.e., cannot be estimated without infor-mation either typically unavailable to the system or expensive to obtain). The manner in which the Type Of Belief of a justified belief is propagated from its justifier beliefs is determined from the Linkage Rule (below) that enables the justification.
The Degree Of Belief is a numerical measure of the validity of beliefs who Type Of Belief is statistical or default. We have not yet explored its u.
A Justification Link contains Linkage Rule and Error Propagation slots (in addition to its links in the justification structure).
The Linkage Rule points to the rule of inference that enables the justified belief to be derived from the justifier beliefs (e.g., modus ponens). It is described in more detail in the next ction.
The Error Propagation specifies the way in which errors in justifier beliefs propagate to the justified belief. This form of this infor-mation for a particular link is inherited from the Linkage Rule for the link.
It allows the Critic to focus on particular justifier beliefs that might be responsible for an error of a specific type in the justified belief.
This slot is filled with a list of the form
((<error-type> <clau>
(<error-type> <clau> <justifier>) ...) ...) For example, suppo the Error Propagation slot for the Justifica-tion Link of the rule Belief17 has the value ((OSLHS Clau1 (OSLHS Clau2 Belief161)). This structure states that if the rule asrted by Belief16 suffers from an OverSpecificLeftHandSide er-ror in Clau2 of its left-hand side then Clau1 of the left-hand side of Belief17 could suffer from the same type of error.
C. LINKAGE RULES
A linkage rule specifies the rationale that allows a justified belief to be derived from its justifier beliefs. It can be either Truth-Prerving or Non-Truth-Prerving. The only type of truth-prerving rule is Deductive. It is esntially a logical proof of the validity of the derivation.
While we hypothesize a number of Non-Truth-Prerving linkage rules, the only one we have ud t
o date is the Abductive rule. Abductive inference allows the conclusion of A→B to be drawn from B→A and A. For example, from the quasi-theoretical state- ment that meningitis caus fever and an obrvation of fever in a patient, abductive inference permits the conclusion that the patient is suffering from meningitis. Becau is commonly associated with a number of other caus as well. the backward (interpretive) form of the causal rule can only be ud to suggest the cau when the manifestation is obrved. (For a general discussion of Abduction, e [7].)
A linkage rule also contains Type of Belief Propagation and Error Propagation slots.
The Type of Belief Propagation slot specifies the way that the Types of Belief of justifier beliefs are propagated to beliefs justified via the rule. Deductive rules, for example, propagate the minimum of Type Of Belief of the justifier beliefs, according to the following partial order: definitional, theoretical, statistical, default. The Ab-ductive rule is a default-producing rule. Regardless of the Type Of Belief associated with the justifier beliefs, default is propagated as the Type Of Belief of the justified belief.
The Error Propagation slot specifies a template for propagating errors from justifier beliefs to justified beliefs.
D. EXAMPLE: THE NFR1 2 JUSTIFICATION
The hand-generated NFR12 justification structure is shown in Figure II-2. Beliefs are indicated by nodes with names Bi. Jus-tification Links are shown between beliefs, but are not explicitly named. A capsule summary of some of the beliefs is shown. We show the entire justification structure to give some feel for its com-plexity. The actual number of nodes should not be taken too riously becau there is considerable latitude available with respect to what constitutes a belief and what constitutes a linkage rule. We have inrted knowledge of the task domain as beliefs and knowledge that is not specific to the domain as linkage rules (e.g., geometry and algebra).3
The general form of the justification is that it is bad on the rela-tive positions of three-dimensional regions associated with a fault. When the regions are penetrated by a borehole, measured data will be related to the zones that are the projections of the regions on the borehole, as manifested in patterns en by a particular tool (e.g., the dipmeter tool).
Nodes that are boxed in the figure are described in detail in the following. Most linkage rules are Deductive. We only note a linkage rule when it has some special significance (e.g., if it is Non-Truth-Prerving). The logic encoding of each belief asrtion is shown. The reader may find it helpful to simply scan the justification at this point and refer back to it while following the example prented in Section A. We only show Error Propagation information that is relevant to the example. The complet
e justifica-tion for NFR12 is found in [11].
B1: If p is a Normal Fault Pattern and there exists a Red Pattern,
p1, such that the length of p1 is less than 50 ft. and p1 is Above the fault plane pattern of p, then p is a Late Fault Pattern.
∀p ∃p1 [[NFP(p) ∧ (RP(p1) ∧ (length p (p1) < 50) ∧
Above p(p1, fpp(p)))] LFP(p)]
Type Of Belief: default
3
Rule justifications provide support for the inferences made in the reasoning trace. Similarly, we might consider construction of another layer of justification – support for the inferences made in linking beliefs in the rule justitica-tion. We have not yet examined this sort of multi-layer justification.
Figure II-2. Justification for NFR12
Linkage Rule: Abductive
Error Propagation: ((OSLHS C2 (OSRHS C2 B2)) ...)
Note: This belief is supported by a non-truth-prerving linkage rule.
B2: If p is a Late Fault Pattern, then p is also a Normal Fault Pattern and there exists a Red Pattern p1 such that the length of p1 is less than 50 ft. and p1 is Above the fault plane pattern of p. ∀p ∃p1 [LFP(p) → [NFP(p) ∧ (RP(p1) ∧ (length p (p1) < 50) ∧
Above p (p1, fpp(p)))] Type Of Belief: default Error Propagation:
((OSRHS C2 (OSRHS C2 B4) (OSRHS C1 B3)) ...) B3: The Distortion Pattern of a pattern is called a Red Pattern ∀p,q [DPofP(p, q) → RP(p)] Type Of Belief: definitional
B4: If p is a Late Fault Pattern, then p is also a Normal Fault Pattern and there exists a p1 such that p1 is the Distortion Pattern of p, the length of p1 is less than 50 ft., and p1 is Above the fault
plane pattern of p.
∀p ∃p1 [LFP(p) → NFP(p) ∧ [DPofP(p1, p) ∧ (length p (p1) < 50)
∧ Above p (p1, fpp(p))]]
Type Of Belief: default Error Propagation:
((OSRHS C2 (OSRHS C1 B6) (OSRHS C1 B7)
(OSRHS C1 B8)) ...)
B6: The length of the distortion pattern of any Late Fault Pattern is less than 50 ft.
∀p [LFP(p) → (length p (dp(p)) < 50)]
Type Of Belief: default Error Propagation:
((OSRHS C1 (OSRHS C1 B65) (OSRHS C1 B72)) (OSRHS C1 (OSRHS C1 B69) (OSRHS C1 B70)
(OSRHS C1 B71)) ...)
B7:∀P [LFP(p) → NFP(p)]
Type Of Belief: definitional
Note: This is supported by beliefs that relate regions to zones and zones to patterns.
通信社B8: If p is a Late Fault Pattern, the distortion pattern of p is Above the fault plane pattern of p.
∀p [LFP(p) → Above p(dp(p),fpp(p))]
Type Of Belief: theoretical
B65: If p is a Late Fault Pattern, the length of the distortion pat. tern of p is equal to the length of the corresponding distortion zone minus the sum of the lengths of the distortion zone's wash-outs, mirror
images, and unprerved bedding zones.
∀p [LFP(p) → length p(dp(p))= length z(dz(z p(p))) −
Iength z(w(dz(z p(p)))) − length z(m(dz(z p(p)))) −
Iength z(ub(dz(z p(p))))
Type Of Belief: theoretical
B69: If p is a Late Fault Pattern, the length of the washout zones of the distortion zone of the zone associated with p is 0.
∀p [LFP(p) → Iength z(w (dz(z p(p)))) = 0)]
Type Of Belief: default
B70: If p is a Late Fault Pattern, the length of the mirror image zones (i.e., zones where the tool is not operating correctly) of the distortion zone of the zone associated with p is 0.
∀p [LFP(p) → (length z(m(dz(z p(p)))) = 0)]
Type Of Belief: default
泡泡宝宝B71: If p is a Late Fault Pattern, the length of the unprerved bedding zones of the distortion zone of the zone associated with p is 0.
∀p [LFP(p) → (length,(u(dz(z p(p)))) = 0)]
Type Of Belief: default
B72: If p is a Late Fault Pattern, the length of the distortion zone of the zone of p is less than 50 ft.
∀p [LFP(p) → (length z(dz(z p(p))) < 50)]
Type Of Belief: statistical
Error Propagation: ((OSRHS C1 (OSRHS C1 859)) ...)
B59: If r is a Late Fault Region, the length of the distortion zone of the zone of r is less than 50 ft.
∀r [LFR(r) → (length z(dz(z r(r))) < 50)]
Type Of Belief: statistical
Error Propagation: ((OSRHS C1 (OSRHS C1 B41)
(OSRHS C1 B44)
(OSRHS C1 B54)) ...)
B41: The thickness of the distortion region of any Late Fault Region is less than 13 ft.
∀r [LFR(r) → (t(dr(r)) < 13)]
Type Of Belief: statistical
Error Propagation: ((OSRHS C1 (OSRHS C1 B36)) ...)
B44: The orientation of the distortion region of every Late Fault Region is less than 75o.
∀r [LFR(r) → (o(dr(r)) < 75)]
Type Of Belief: statistical
Error Propagation: ((OSRHS C1 (OSRHS C1 B42)) ...)
B42: The orientation of every Late Fault Region is less than 75°.
∀r [LFR(r) → (o(r) < 75)]Type Of Belief: statistical
Error Propagation: ((OSRHS C1 (OSRHS C1 B37)) ...)
B54: If r is a Late Fault Region, the length of the zone of the distortion region of r is equal to the thickness of the distortion region of r divided by the cosine of the orientation of the distortion region.
∀r [LFR(r) → (length z(z r(dr(r))) = t(dr(r))/cos(o(dr(r))))]
Type Of Belief: definitional
B36: If r is a Late Fault Region in the Gulf Coast, the thickness of r is less than 13 ft.
∀r [LFR(r) ∧ G(r) → (t(dr(r)) < 13)]
Type Of Belief: statistical
Error Propagation: ((OSRHS C1 (UE B75) (SPE B78)) ...)
Note: Either the threshold (13 ft.) or the comparator (<) could be incorrect.
B39: ∀r [R(r) ∧ (G(r) → (t(dr(r)) < 13))] → (t(dr(r)) < 10)
Type Of Belief: definitional
Note: This is an instantiation of P in B38 to be the distortion region thickness condition.
B40: ∀r [R(r) ∧ (G(r) → (o(r) < 75))] → (o(r) < 75)
Type Of Belief: definitional
Note: As in B39 with P the orientation condition.
B37: If r is a Late Fault Region in the Gulf Coast, its orientation is less than 75°.
∀r [LFR(r) ∧ G(r) → (o(r) < 75)]
Type Of Belief: statistical
Error Propagation: ((OSRHS C1 (UE B77) (SPE B79)) ...)
Note: Analogous to B36.
B38: For all regions r and predicates P, if r being in the Gulf Coast implies P(r), then P(r).
∀r,P [R(r) ∧ (G(r) → P(r))] → P(r)
Type Of Belief: definitional
Note: This is a cond order axiom. It is ud to eliminate the Gulf Coast precondition. This precondition is not carried over into NFR12 becau it rarely fails, and when it does fail, the effect is not disastrous. The justification records the way in which it is important (i.e., through the default values for distortion region thickness and fault orientation).
B74: The thickness of the distortion region of a Late Fault Region in the Gulf Coast has relation R1to the parameter τ.
∀r [LFR(r) ∧ G(r) → R1(t(dr(r)), τ)]
Type Of Belief: definitional
B75: τ = 13
Type Of Belief: statistical
B78: R1 = <
Type Of Belief: theoretical
付款条件B76: The orientation of a Late Fault Region in the Gulf Coast has the relation R2 to the parameter θ.
∀r [LFR(r) ∧ G(r) → R2(o(r), θ)]
Type Of Belief: definitional
B77:θ= 75
Type Of Belief: statistical
B79: R2 = <
Type Of Belief: theoretical
III. REASONING FROM ERROR TYPES
AND JUSTIFICATION STRUCTURES
The Critic begins by considering a specific instance in which the ur corrects or augments the Performance Program's conclu-sions. Given any such failure, the first step is to determine specific types of errors in specific rules. that could have produced this failure. For example, if the ur deletes a system-generated con-clusion, then the rule that suggested this conclusion may be suspected to be in error4, with the possible error types Over-GeneralLeftHandSide and OverSpecificRightHandSide (either of the error types could have led this rule to suggest the incorrect conclusion). Similarly, if the ur adds a conclusion, then tho rules who right-hand side mentions that conclusion but which did not trigger are identified as possible errors of type Over-SpecificLeftHandSide, while rules that did apply but did not make the indicated conclusion may be suspected of OverGeneralRight-HandSide errors. (Of cour, another plausible explanation for a failure of this type is abnce of an appropriate rule.)
At this point, given a suspected belief (e.g., a rule), and a list of possible error types for the belief, the Critic examines the justifica-tion for that belief in order to generate a list of candidate hypothes regarding possible caus of the error (i.e., bugs in the supporting beliefs, or approximations in the justification links relating the belief to its supporting beliefs). The method for generating and pruning hypothes about the cau of the error is summarized below:
Explain(possible-errors, belief)
begin
<for each possible error, enumerate supporting beliefs that could have caud the error,
along with their suspected error type> <prune the suspects>
<attempt to determine the correctness of each suspect in the current situation, removing
tho shown to be correct.>
<rank remaining suspects according to their
Type Of Belief, and remove suspects who
Type Of Belief is "definitional".> <recursively Explain each remaining suspect>
end
This method for identifying suspect beliefs makes u of veral kinds of information recorded in the
漩涡香磷
justification structure for the offending rule, as well as knowledge of the context in which the rule failure occurred. The kinds of information and their u may-be summarized as follows:
a股什么意思The logical dependencies of one belief on another, recorded in the justification structure, provide the basis for generating can- didate suspects. The justification structure is the basis for a "complete" generation of suspects. in the n that the only possible errors in the underlying domain knowledge that could produce the detected rule error are tho involving beliefs men-tioned in the rule's justification structure. All the other sources of knowledge that enter into the process rve to constrain the suspects generated on this basis.
4
带狗的成语Of cour it is possible that this rule is correct, that it should not have been applied, but that its preconditions were satisfied only becau of an error in an earlier rule. Thus, in general there will be veral alternative rule suspects. The Critic weighs the alternative suspect rules by examining their justifications, and determining the plausible caus of errors for each. The additional knowledge about error types, together with the Error Propagation knowledge associated with each justification link, allows the Critic to prune the t of suspects. In other words, certain supporting beliefs on whic
h the offending rule is bad can be pruned as suspects when it can be shown that no error in tho beliefs could produce the detected rule error.
Knowledge of the situation in which the failure occurred may allow the system to parately verify the correctness of suspected supporting beliefs. For example, rule NFR12 depends, among other things, upon the assumption that normal faults have associated distortion regions. If this rule is suspected of a failure, that supporting assumption may become suspect. But if there is strong direct evidence in the current situation that a distortion region is prent, then that suspect may be pruned.
Type Of Belief knowledge propagated from supporting beliefs as specified by the associated linkage rule allows further ranking and pruning of suspects.
A. EXAMPLE: USE OF THE NFR12 JUSTIFICATION
As an example, we assume a scenario in which the ur has in-dicated that a particular hypothesized "normal fault" should be specialized to a "late fault." The LAS takes this as an indication that it has committed an error of omission by failing to make this specialization on its own, and invokes the Critic to explain the failure.
The Critic begins by examining the reasoning trace of the Perfor-mance Program and determines that rule NFR12 could have drawn the correct conclusion, but failed to match. For simplicity we will assume that NFR12 is the only rule that could have drawn the correct conclusion5 The Critic examines the situation in which the rule was attempted and, through interaction with the ur determines that: (i) the ur agrees that the normal fault con- clusion drawn by the Performance Program is correct; and (ii) the Performance Program detected a satisfactory red pattern that is above the fault plane pattern of the normal fault. but is longer than 50 ft. In general, the Critic us a combination of the reasoning trace and the rule justifications to track down the source of the failure. If, for example, the red pattern had been found to be unacceptable to the ur, then attention would have been focusd on the detector for that pattern.
As a result of this preliminary analysis, the Critic hypothesizes that rule NFR12 has committed an error of type OverSpecificLeftHand-Side, specifically in clau 2. The Critic now attempts to explain this error in terms of the justification for NFR12 (i.e., for B1 in Figure II-2), in order to determine which types of errors in its sup-porting beliefs could have caud this error. The following paragraphs summarize the generation and pruning of suspect beliefs generated by the procedure Explain described above.
In the first step of this procedure, the justification of B1 is ex- amined to find that it depends upon B2, and that the OverSpecific-LeftHandSide error in B1 can only be explained by an Over-
5
Of cour if other rules are available which could have drawn the correct conclusion, then their justifications must be examined to generate additional hypothes. While this may add substantially to the number of hypothes that the Critic must consider, it does not change the reasoning that the Critic goes through in considering each hypothesis.
