View
214
Download
0
Tags:
Embed Size (px)
Citation preview
Luigi Portinale, Pietro Torasso and Diego Margo
Selecting Most Adaptable Diagnostic Solutions through Pivoting-Based Retrieval
Teacher : C.S. Ho
Student : L.W. Pan
No. : M8702048
Date : 10/1/99
1999/10/1 Li-we Pan 2
Why need more retrieval
• Before: aim at highest similarity(surface feature)
• Now : adaptation estimation & adaptation effort(trade off)– Can prune non-adaptable or hard to adapt cases– An approach to adaptation-guided retrieval
based on a tight integration between adaptation effort estimation and retrieval of past diagnostic solutions
1999/10/1 Li-we Pan 3
On the Adaptation of Diagnoses• ADAPtER : a diagnostic system integrating a formal
theory of model-based diagnosis with CBR.• Def1: a diagnostic problem is a tuple
– DP = <<T,H>,CXT,<Ψ+,Ψ->>• T : a set of logical formulae representing the
behavioral model of the system to be diagnosed• H : a set of diagnostic hypotheses• CXT : the set of contextual information of the problem• Ψ+,Ψ-: the set of manifestation to be accounted for
(covered)
1999/10/1 Li-we Pan 4
Cont.– OBS:the set of observed manifestations’
• Ψ+OBS, Ψ-= {m(a)|m(b)OBS,b≠a}
– MANA:abnormal manifestations– MANN:normal manifestations
• Ψ+=OBSA,OBSA=MANA∩OBS
• Def2:– DP = <<T,H>,CXT,<Ψ+,Ψ->>– A diagnosis a set EH m(a)Ψ+ T CXT E ├ m(a); m(a)Ψ- T CXT E ├ m(a);
1999/10/1 Li-we Pan 5
Estimating Adaptation• Stored case is represented as the tuple
C=<CXTall, CXTsome, OSB, SOL>– CXTall:the set of contexts relevant to every solutions of the
cases;– CXTsome:the set of contexts relevant to some(but not all)
solutions of the cases;– OBS:the set of manifestations observed in the case– SOL=<<H1,EXPL(H1,CXT1,)>,…,< Hn, EXPL(Hn ,CXTn)>>
is the list of solutions• Hj : a set of diagnostic hypotheses• CXTj : the set of context relevant to the j-th solution• EXPL() : the derivational trace form Hj and CXTj observable features
1999/10/1 Li-we Pan 6
How to estimate• Input case: CI =<CXTI,OBSI>• Retrieval solution Sj = <Hj,EXPL(Hj,CXTj)>• (compare CI and Sj)
– Compare CXTI with CXTj
– Manifestations in OBSI with those in EXPL(Hj,CXTj)
• Context : Slightly or totally incompatible • Manifestation :
1. input case m(a) & retrieval solutions m(b) has a different value
2. Only input case m(a) has value
1999/10/1 Li-we Pan 7
Heuristic estimate• Let :
– ρ: the estimated cost of inconsistency removal– γ: that of explanation construction
1. αCONFLICT(m(a)) = – ρ +γ if m(a) to be covered and m(b) supported– γ if m(a) to be covered and m(b) not supported– ρ if m(a) not to be covered and m(b) supported– 0 otherwise
2. αNEW(m(a)) = – γ if m(a) to be covered– 0 otherwise
• h(Sj) =ΣαCONFLICT(m(a))+ΣαNEW(m(a))+δ|SI(Sj)|– SI(Sj) : the set of contexts of solution Sj slightly incompatible with CXTI– δ: the adaptation weight assigned to them
1999/10/1 Li-we Pan 8
The PBR Algorithm• Input : a case C1 = <CXTI,OBSI>
• Output : a set of solutions Sj = <Hn,EXPL(Hn,CXTn)> with minimal h(Sj)
1. Filtering. Construct a first set CC1 of candidate cases by following indices
• Only cases having at least one feature in common with the input case
2. Context-Based Pruning. Restrict the set CC1 into the set CC2 by removing each case C such that there is a context in CXTall totally incompatible with a context in CXTI
• Rejecting cases having in all their solutions contextual information conflicting with the input one
1999/10/1 Li-we Pan 9
Cont.
3. Bound Computation. For every case C CC2 compute
a pair [hlC, hu
C], SjSOL hlC <= h(Sj)<= hu
C
– Computations of bounds on the adaptation estimates of solutions of cases
4. Bound-Based Pruning. Restrict CC2 to CC3 by
removing every case C such that hlC>a, a = minchu
C
– Reject cases which have definitively no solutions with minimal estimate
1999/10/1 Li-we Pan 10
Cont.5. Pivoting. (…)
– No deep investigations on the solutions of the case is performed
1999/10/1 Li-we Pan 13
Conclusion
• Simple memory organization avoiding the space problems of more complex organizations like E-MOP
• Allow one to obtain the best possible accuracy in terms of adaptation effort estimate
• Retrieval time is considerably reduced by the combination of pivoting and pruning techniques
1999/10/1 Li-we Pan 14
program• Utility : the match rate(hit features/total feature)• EU : ΣP x EUnext
• P : ? (domain similarity)• Adaptation knowledge :
– If (query value –case’s value) / case’s value >= 95%– Then can adaptability– Else cannot adaptability
• Adapt method : replace• Question : each input case(query) need rebuild the
tree?