1. Béatice Fuchs, Jean Lieber, Alain Mille and Amedeo Napoli ICCBR 99 Towards a Unified Theory of Adaptation in Case-Based Reasoning Teacher : Dr. C.S. Ho Student : L.W. Pan No. : M8702048 Date : 11/26/99

2. 1999/11/26Li-we Pan2 Basic step in CBR 1.Retrieval : search similar problem in the case base 2.Adaptation : adapt the similar problem 3.Possible memorization of the problem and its solution, in the perspective of a future reuse Combine the retrieval and adaptation steps in a unique planning process

3. 1999/11/26Li-we Pan3 Adaptation scheme Retrieval of a similar case aims at finding an adaptable case target Target problem source source problem in CB Retrieval process Sol(source)Sol(target)? Adaptation process

4. 1999/11/26Li-we Pan4 Approach in Adaptation Adaptation-guided retrieval [smyth & keane, 1996] Adaptation rely on a search in memory. based on contextual similarity. [Leake] Seen as a constraint satisfaction problem(CSP) The framework of hierarchical adaptation [ECAI96] Case-based planning is presented in the framework of description logics and is analyzed form a formal viewpoint [koehler, 1996] –idx(init target )  idx(init source ) and/or idx(goal source )  idx(goal target ) –And : strong retrieval : each source cases verifying the above conditions can be reused in order to solve target –Or : week retrieval : best candidate case is the one sharing the more sub-goals or preconditions with the problem target

5. 1999/11/26Li-we Pan5 Two challenges 1.Be able to define generally what is a case & to model adaptation independently of any application domain 2.Combine the retrieval and adaptation operations in a single operation The adaptation of a case is considered as a plan adaptation Adaptation which has involved the notions of reformulation and similarity path –Reformulation : an order pair (r P, r S ) –r P : relation : pb 0 r P pb 1 –r S : Sol(pb 0 ) -> Sol(pb 1 ) –Similarity path : pb 0 r S pb 1 … pb n

6. 1999/11/26Li-we Pan6 A case as a plan

7. 1999/11/26Li-we Pan7 Plan Adaptation Generative approach –All the knowledge necessary for generating a solution plan form scratch Transformational approach –The previous hypothesis is not necessary and that is based on reuse Decomposed –A generalization of the source case compatible with the target problem –A specialization satisfying the descriptors of the target problem

8. 1999/11/26Li-we Pan8 Adapting a case like a plan Solving a problem consists in building the list of the solution descriptors leading to the satisfaction of the goal statement source Sol(source) idx(source) Sol(idx(source)) idx(target) Sol(idx(target)) target Sol(target)

9. 1999/11/26Li-we Pan9 Example (route) TNeighbor(T) A{B, E, G} B{A, C, E} C{B, D, E} D{C, F, I} E{A, C, F, G} F{D, E, H} G{A, E, H} H{F, G} I{D} D(T,S) = 0, if T=S = 1, if T  neighbors(S) =+∞, if T  neighbors(S) [2] idx(target) = (init target, goal target ) idx(source) = (init source, goal source ) [3] dict(idx(source),idx(target) =[3] dict((init source,goal source ), (init target, goal target )= d(init source, init target )+d(goal source, goal target ) [4]

10. 1999/11/26Li-we Pan10 Case Base Case noidx(source)Sol(source) 1(A, C){((A,B),(B,C)), 2:00} 2(C, G){((C,E),(E,G)), 1:30} 3(A, D){((A,B),(B,C),(C,D)), 3:00} 4(D, C){((D,C)), 1:00} 5(H, E){((H,F),(F,E), 2:00} 6(E, B){((E,C),(C,B), 2:15} 7(G, D){((G,E),(E,C),(C,D), 2:45} 8(B, I){((B,C),(C,D),(D,I), 3:15}

11. 1999/11/26Li-we Pan11 Distance table ([1] [2] [4]) Case noidx(source)dist(idx(source), idx(target)) 1(A, C)2 2(C, G)+∞ 3(A, D)1 4(D, C)+∞ 5(H, E)+∞ 6(E, B)+∞ 7(G, D)+∞ 8(B, I)1

12. 1999/11/26Li-we Pan12 A case is a kind of plan Problem = initial context + goal statement (descriptors which be satisfied by the solution) Descriptor = (attribute, value) or constraint Solution = Σdescriptors A plan is considered as a triple (I, G, O) –I: initial state –G: goal state –O : a set of operations allowing to satisfy the goal statement given the data associated with O

13. 1999/11/26Li-we Pan13 Solving the target problem source Sol(source) idx(source) Sol(idx(source)) idx(target) Sol(idx(target)) target Sol(target) (A,D, 10/6/99, Fiat Uno) (A, D) (B, D) (B, D, 15/2/99, 2CV, Citroen) {((A,B),(B,C),(C,D),3:00)} ((A,B),(B,C),(C,D))((B,C),(C,D)) {((B,C),(C,D)),1:45} Select (A,D), (B,I) Remove(A,B)

14. 1999/11/26Li-we Pan14 Discussion This model is simple and clear puts in question the common idea saying that adaptation can be only contextual Be used for a deep discussion about the two key phases of the CBR cycle Adaptation can be considered as a sequence of generalizations, specializations and substitutions performed at different levels

15. 1999/11/26Li-we Pan15 Future Work Justify more precisely the assumption mode on the type of problems and show that it works correctly on every type of problems Make more progress in the formal description of this work in order to formalize the retrieval-adaptation process indepdently of any context The learning step of case-based reasoning must be integrated