1 / 29

Preferential Defeasibility: Utility in Defeasible Logic Programming

Preferential Defeasibility: Utility in Defeasible Logic Programming. Fernando A. Tohmé Dept. of Economics Guillermo R. Simari Dept. of Computer Science and Engineering U NIVERSIDAD N ACIONAL DEL S UR ARGENTINA. Outline. Motivation The Argumentation Framework Comparison Criteria

vila
Download Presentation

Preferential Defeasibility: Utility in Defeasible Logic Programming

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Preferential Defeasibility: Utility in Defeasible Logic Programming Fernando A. Tohmé Dept. of Economics Guillermo R. Simari Dept. of Computer Science and Engineering UNIVERSIDAD NACIONAL DEL SUR ARGENTINA

  2. Outline • Motivation • The Argumentation Framework • Comparison Criteria • Example and Results • Conclusions

  3. Deafeasible Logic Programming: DeLP Strict Rules  Facts Defeasible Rules  A Defeasible Logic Program (dlp) is a set of facts, strict and defeasible rules denoted  = (, ) bird(X) chicken(X) chicken(tina) bird (X) penguin(X) penguin(opus) flies(X) penguin(X) scared(tina) flies(X)  bird(X) flies(X) chicken(X) flies(X) chicken (X), scared(X)

  4. Defeasible Argumentation Def: Let L be a literal and (, ) be a program. , L is an argument, for L, if  is a set of rules in  such that: • There exists a defeasible derivation of L from   ; • The set    is non contradictory; and •  is minimal, that is, there is no proper subset  of such that satisfies 1) and 2).

  5. buy_shares(X) good_price(X) buy_shares (X) good_price(X), risky(X) risky(X)  in_fusion(X, Y) risky(X) in_debt(X) risky(X) in_fusion(X, Y), strong(Y) good_price(acme) in_fusion(acme, estron) strong(estron) buy_shares(acme) good_price(acme) risky(acme) good_price(acme) in_fusion(acme, enron) in_fusion(acme, enron) {buy_shares(acme)  good_price(acme), risky(acme)., risky(acme)  in_fusion(acme, enron).}, buy_shares(acme)

  6. buy_shares(acme) good_price(acme) risky(acme) good_price(acme) in_fusion(acme, enron) in_fusion(acme, enron) , Q is a subargument of, L if is an argument forQ and   = {risky(acme)  in_fusion(acme, enron). }  = {buy_shares(acme) good_price(acme), risky(acme)., risky(acme)  in_fusion(acme, enron). }

  7. Counter-argument buy_shares(acme) good_price(acme) risky(acme) risky(acme) good_price(acme) in_fusion(acme,estron) in_fusion(acme,estron) strong(estron) in_fusion(acme,estron) in_fusion(acme,estron) strong(estron)   { risky(acme), risky(acme)} is a contradictory set

  8. Argument Comparison: Generalized Specificity Def:Let = (, ) be a program, let G be the set of strict rules in  and let F be the set of all literals that can be defeasibly derived from . Let 1, L1and 2, L2be two arguments built from , where L1, L2F. Then 1, L1isstrictly more specific than2, L2if: • For all H F, if there exists a defeasible derivation GH  1L1whileGHL1then GH  1L2, and • There existsH F such that there exists a defeasible derivation GH 2 L2andGH L2 but GH 1 L1 (Poole, David L. (1985). On the Comparison of Theories: Preferring the Most Specific Explanation. pages 144—147 Proceedings of 9th IJCAI.)

  9. Defeaters P L Q    An argument , Pis a defeater for , L if , P is a counter-argument , L that atacks a subargument , Q de , L and one of the following conditions holds: (a), Pis better than, Q (proper defeater), or (b), Pis not comparable to, Q (blocking defeater)

  10. Argumentation Line L4 L1 L2 L3 L0 1 2 3 4 0 Given = (, ), and 0, L0an argument obtained from. An argumentation linefor0, L0is a sequence of arguments obtained from , denoted  = [0, L0, 1, L1, …] where each element in the sequence i, hi, i > 0is a defeater fori-1, hi-1.

  11. Argumentation Line L1 L4 L2 L3 L0 1 2 3 4 0 Given an argumentation line = [0, L0, 1, L1, …], the subsequence S= [0, L0, 2, L2, …]contains supporting arguments and I= [1, L1, 3, L3, …]are interfering arguments. S

  12. Argumentation Line L1 L4 L2 L3 L0 1 2 3 4 0 Given an argumentation line = [0, L0, 1, L1, …], the subsequence S= [0, L0, 2, L2, …]contains supporting arguments and I= [1, L1, 3, L3, …]are interfering arguments. I

  13. Acceptable Argumentation Line Given a program= (, ), an argumentation line  = [0, L0, 1, L1, …] will beacceptable if: •  is a finite sequence. • The sets Sof supporting arguments is concordant, and the set Iof interfering arguments is concordant. • There is no argument k, Lk in  that is a subargument of a preceeding argument i, Li,i < k. • For all i, such that i, Liis a blocking defeater for i-1, Li-1, if there exists i+1, Li+1 then i+1, Li+1is a proper defeater for , Li (i.e.,, Li could not be blocked).

  14. 0 1 2 4 4 5 3 2 1 3 4 2 Dialectical Tree Given a program= (, ), a literal Lwill be warranted if there is an argument , L built from , and that argument has a dialectical tree whose root node is marked U. That is, argument , L is an argument for which all the possible defeaters have been defeated. We will say that is a warrant for L. 2 3 1 , L 3

  15. U D D D U U U U D D U U Marking of a Dialectical Tree *, L

  16. Answers in DeLP Given a program= (, ), and a query forL the posible answers are: • YES, if L is warranted. • NO, if L is warranted. • UNDECIDED, if neither Lnor L are warranted. • UNKNOWN,ifL is not in the language of the program.

  17. A Comparison Criterion • A key element for the warrant procedure is the defeat relation. • Generalized specificity is a purely syntactic comparison criterion and it is introduced as a choice among other possible comparison criteria for comparing arguments. • Here, we will offer an extension of generalized specificity using pragmatic considerations.

  18. A Comparison Criterion • We will allow utility values for facts and rules. • Decision-Theoretic Defeasible Logic Programming will be represented as = (, , , B), where  and  are as before, B is a Boolean algebra with top  and bottom, and  is defined :     B. • Band ()are used to represent the explicit preferences of the user in the sense that given two pieces of information 1, 2 in   , if 1 is strictly more preferred than 2 then (1) B(2) where Bis the order of B. • The elements  of    which are most preferred receive the label () = . • From the preferences over   , we can find preferential valuesover defeasible derivations.

  19. A Comparison Criterion • Given a defeasible derivation of Lfrom   , L1, L2, …,Ln, let D be the set { L1, L2, …, Ln } and {1, 2, …, n } a set such that iyieldsLi .Then, that derivation yields for its conclusion L a valueV(L, )i=1..nV(Li, i). Inductively: • V(L, ) () if Lis a fact, or • V(L, )  () k=1..mV(Bk, k)if  is a rule with head Land body B1, B2, …,Bm and k is a rule used to derive Bk. • The intuition is that a conclusion is as strongly preferred as the weakest of either its premises or the rule used in the derivation.

  20. A Comparison Criterion • By extension, an argument , L gives a value for its conclusion V(L, )  DV(L, D), where D is a derivation that uses all the defeasible rules in A and only those defeasible rules. • Note that there could be many different derivations D that contain the defeasible rules in . • In that manner, DV(L, D), will obtain the lowest value among the defeasible derivations of Lthat use the defeasible rules in .

  21. A Comparison Criterion • Let F be the set of all literals that can have a defeasible derivation from   . Any subset H F be has a value V(H) L HDV(L, D) • This means that H is as valuable as the most valuable of its elements, which in turn is as valuable as the weakest of its derivations. • We can use this notion to redefine specificity obtaining a relation of preferential specificity.

  22. Preferential Comparison Def:Let = (, , , B) be a program, let G be the set of strict rules in  and let F be the set of all literals that can be defeasible derived from . Let 1, L1and 2, L2be two arguments built from , where L1, L2F. Then 1, L1isstrictly more preferentialy specific than2, L2if: • For all H F, if there exists a defeasible derivation GH  1L1whileGH L1then GH  1L2, and • There existsH F such there exists a defeasible derivation GH 2L2andGH  L2 but GH 1L1 • For evey H verifying (1) andHverifying (2) holds V(H) BV(H)

  23. Example • Consider a classical example in defeasible argumentation where preferences are defined for B = { 0, 1 }, with 0  1:  { bird(X) penguin(X) (1), penguin(tweety) (0), bird(tweety) (1) }  { flies(X)  penguin(X) (1), flies(X) bird(X) (1) } • Notice that bird(tweety) yields two values: V(bird(tweety), {penguin(tweety), bird(tweety)})  min(0,1)  0 and V(bird(tweety), )  1, because the fact thattweety is a penguinhas a preference of 0 while the rule used to derive that it is a bird has a preference of 1.

  24. Example Now, consider the two arguments: {flies(X)  penguin(X)}, flies(X) and {flies(X)  bird(X)}, flies(X) then if we consider H  { penguin(tweety) } and H { bird(tweety) } we have that H { bird(X)  penguin(X) }  flies(tweety), but H { bird(X)  penguin(X) }  {flies(X)  penguin(X) } flies(tweety) H { bird(X)  penguin(X) }  { flies(X) bird(X) } flies(tweety) On the other hand H { bird(X)  penguin(X) }  flies(tweety), but H { bird(X)  penguin(X) }  { flies(X) bird(X) } flies(tweety) H { bird(X)  penguin(X) }  {flies(X) penguin(X) } flies(tweety) Therefore { flies(X)  penguin(X) }, flies(X) is strictly more specific than { flies(X)  bird(X) }, flies(X)

  25. Example • We found that { flies(X)  penguin(X). }, flies(X) is strictly more specific than { flies(X)  bird(X). }, flies(X) but is not strictly more preferentially specific since we have that V(H )  max(V(bird(tweety), ), V(bird(tweety), {penguin(tweety), bird(tweety)})  max(1, 0)  1 while V(H) V( penguin(tweety), )  0

  26. Results Proposition: If 1, L1is strictly more preferentially specific than2, L2 then 1, L1is strictly more specific than2, L2. Proposition: The relation strictly-more-preferentially-specific-than in program  (, , , B) is equivalent (i.e., yields the same subset of ARGARG where ARG is the class of argument structures) to the relation strictly-more-specific-than in program  (, ) if and only if for every pair of argument structures 1, L1, 2, L2 ARG, 1, L1 is strictly-more-specific-than2, L2 and for every pair of their corresponding activation sets H, HF, V(H) BV(H) .

  27. Results Proposition: Given a query Q in the preferential defeasible logic program  (, , , B), and an argument structure A, Q, its tagged dialectical tree is identical to T*A, Qin  (, ) iff the relation strictly-more-preferentially-specific-than for program  is equivalent to the relation strictly-more-specific-than in program  over ARGQ, where ARGQ is the class of all arguments that are either labels of the dialectical tree TA, Q or subarguments of them.

  28. Results Corollary: Given a query Q and an argument structure A, Q, the answer to Q in the preferential defeasible logic program  (, , , B) is identical to its answer in  (, ) iff the relation strictly-more-preferentially-specific-than for  is equivalent to the relation strictly-more-specific-than in  over ARGQ.

  29. Questions?

More Related