Preferential defeasibility: utility in defeasible logic programming

The development of Logic Programming and Defeasible Argumentation lead to Defeasible Logic Programming. Its core resides in the characterization of the warrant procedure. Defeasible Argumentation has provided a solid foundation over which the standard formalization of this procedure has been constructed. A key element in the warrant procedure is the criterion according to which two contradicting arguments are compared and eventually one of them deemed as defeating the other. The purely syntactic Specificity criterion has constituted the main choice in the design of the warrant procedure. Nevertheless, it seems unreasonable to limit the possibilities of comparison among arguments only to syntactic criteria. The justification of the methods of Defeasible Argumentation are largely pragmatic. Therefore, it seems sensible to expand the set of comparison criteria to incorporate other pragmatic reasons for choosing one argument over another. Decision Theory is the natural choice to model decision-makers. Clearly, as a discipline, it has characterized and introduced formal models in all kinds of pragmatic criteria used in actual choice situations. Here, we will present the framework of Preferential Defeasible Logic Programming. This framework extends the original comparison criteria of specificity redefining it by allowing different preferential values for activation sets. This extension leads to interesting results where the decision is taken considering not only specificity, but also the corresponding pragmatic relation of preferences. Introduction and Motivation The development of defeasible reasoning in the last decades (Pollock 1987; Simari & Loui 1992; Nute 1994; Pollock 1995; Chesñevar, Maguitman, & Loui 2000), lead to the creation of an alternative form of declarative programming, Defeasible Logic Programming ∗Partially supported by Secretaŕıa General de Ciencia y Tecnoloǵıa de la Universidad Nacional del Sur, the Agencia Nacional de Promoción Cient́ıfica y Tecnológica (PICT 2002 Nro 13096) and the National Research Council (CONICET), Argentina (DeLP) (Garćıa 2000; Garćıa & Simari 2004). This formalism blends Logic Programming with Defeasible Argumentation, allowing the representation of tentative knowledge and leaving for the inference mechanism the task of finding the conclusions that the knowledge base warrants (Chesñevar et al. 2003). DeLP inherits from Logic Programming (LP) the formal characterization of programs as sets of rules. The difference is that DeLP considers two kinds of rules. On one hand, strict rules, which are assumed to represent sound knowledge and are handled as the rules in LP. On the other hand, defeasible rules represent tentative knowledge that may be defeated by other information. Again as in LP, DeLP operates by answering queries posed by the users. A query Q succeeds if there exists a warranted argument A for Q. Arguments are constructed using both types of rules and facts (which can be seen as special cases of strict rules). The inference mechanism generates all the arguments that either support or contradict Q. Then, it runs a warrant procedure that determines which arguments end up undefeated. If there exists at least one argument warranting Q, it yields a positive answer. The core of DeLP resides in the characterization of the warrant procedure. Defeasible Argumentation has provided a solid foundation over which the standard formalization of this procedure has been constructed. A key element in the warrant procedure is the criteria according to which two contradicting arguments are compared and eventually one of them deemed as defeating the other (Simari, Chesñevar, & Garćıa 1994). Pure syntactic criteria like specificity are both easy to understand and to implement, and therefore constituted the main choice in the design of the warrant procedure (Poole 1985; Simari & Loui 1992; Stolzenburg et al. 2003). However, it seems unreasonable to limit the possibilities of comparison among arguments only to syntactic criteria. The justification of the methods of Defeasible Argumentation are largely pragmatic, that is, based on how human reasoners perform in the actual world. That is why the ultimate test for systems of defeasible reasoning is how they respond to certain benchmark problems. Therefore, it seems sensible to expand the criteria of comparison to incorporate other pragmatic reasons for choosing one argument over another. Since the issue is one of how to make choices, it is natural to resort to the tools of Decision Theory (DT). In fact, DT’s goal is to capture in formal models the actual behavior of decision-makers. Therefore, as a discipline, it has characterized and introduced in formal models all kinds of pragmatic criteria used in actual choice situations (Loui 1998). Any analysis in Decision Theory begins with the characterization of the preferences of a decision-maker, in the form of an ordering of the alternatives, one of which she has to select. Once this order is defined, the rational behavior of the agent is to choose one of the maximal elements. To obtain sound outcomes it is required that the ordering of alternatives be complete and transitive. This so called weak order can be easily represented by means of a utility function, that assigns to each alternative its rank (or a monotone transformation of the rank). In other terms, each alternative receives a numerical tag and the alternative with the highest value is chosen (Doyle 1990; Osborne & Rubinstein 1994). To introduce decision-theoretic tools into DeLP we have to define what the alternatives are in this case. A hasty answer could be the arguments that can be constructed, but this is not an appropriate option. The idea of introducing this tools in DeLP is to preserve its declarative nature but expand the possible set of warrants for queries. Since the generation of arguments is the task of the inference engine, to assign utilities to them involves to reintroduce the user in a process that should remain opaque for her (Loui 1990; Tohmé 2002). Therefore, the preferences must be already defined in the characterization of a Defeasible Logic Program. That leaves only two possibilities: utilities must be attached either to rules or to facts. There are no major reasons to prefer one possibility over the other, since preferences are to be defined by the user, who may find reasons to rank both the rules and the facts to be used in the construction of arguments. Notice that if we allow utilities to be attached to rules, the distinction between strict and defeasible fades away, since every rule becomes defeasible just because it can be outranked by another rule. Again, this is a decision to be left to the user. The plan of the rest of this paper is as follows. In section 2 we will present the rudiments of DeLP without utilities. In section 3 we introduce utilities and describe how arguments may become ranked by the inference engine. Section 4 discusses possible extensions for this work. Defeasible Logic Programming In order to discuss the introduction of utilities in DeLP, we have to present the basics of this formalism (see (Garćıa & Simari 2004) for a full presentation). It has a language with three disjoint components: • Facts, which are ground literals representing atomic information (or the negation of atomic information). • Strict Rules of the form L0← L1, . . . , Ln, where L0 is the head and {Li}i>0 is the body. Each Li in the body or the head is a literal. • Defeasible Rules of the form L0 –≺L1, . . . , Ln, where L0 is the head and {Li}i>0 is the body. Each Li in the body or the head is a literal. Then, a Defeasible Logic Program P is a set of facts, strict rules, and defeasible rules. P = (Π,∆), where Π denotes the set of facts and strict rules, while ∆ denotes the set of defeasible rules. For each query Q there are four possible answers: YES, NO, UNDECIDED or UNKNOWN. To determine which answer is correct, we need the notion of argument. Given a program P = (Π,∆) and a literal L, 〈A, L〉 is an argument structure for L. A is a set of defeasible rules in ∆ such that: 1. there exists a defeasible derivation of L from Π ∪ A. That is, there exists a finite sequence L1, . . . , Ln = L of ground literals, such that each Li is either a fact in Π or there exists a rule in Π ∪ A with Li as its head, and every literal Bj in the body is such that Bj ∈ {Lk}k 0 〈Ai+1, Li+1〉 is a defeater of 〈Ai, Li〉. ΛS = [〈A0, L0〉, 〈A2, L2〉, 〈A4, L4〉, · · · ] is the sequence of supporting argument structures of Λ, while the sequence of interfering ones is ΛI = [〈A1, L1〉, 〈A3, L3〉, 〈A5, L5〉, · · · ]. An acceptable argumentation line in a defeasible program P = (Π,∆) is a finite sequence Λ = [〈A0, L0〉, · · · , 〈An, Ln〉] such that: 1. Both ΛS and ΛI are concordant, i.e., there is no P such that both P and ¬P have defeasible derivations from Π∪ ⋃bn 2 c i=0 A2i and no P ′ with defeasible derivations for both P ′ and ¬P ′ from Π ∪ ⋃bn−1 2 c i=0 A2i+1. 2. No argument 〈Ak, Lk〉 ∈ Λ is a subargument of an argument 〈Aj , Lj〉, i.e., Ak 6⊂ Aj , for j < k. 3. For each i < n, if 〈Ai, Li〉 is a blocking defeater of 〈Ai−1, Li−1〉 then 〈Ai+1, Li+1〉 is a proper defeater of 〈Ai, Li〉. To answer a query Q, the warrant procedure builds up a candidate argument structure 〈A, Q〉. Then, it associates to this argument a dialectical tree T〈A,Q〉 as follows: 1. The root of the tree is labeled, 〈A0, Q0〉, i.e., A0 = A and Q0 = Q. 2. Let n be a non-root node, with label 〈An, Qn〉 and Λ = [〈A0, Q0〉, · · · , 〈An, Qn〉] the labels in the path from the root to n. Let B = {〈B1,H1〉, · · · , 〈Bk,Hk〉} be the set of all the defeaters for 〈An, Qn〉. For 1 ≤ i ≤ k, if Λ ′ = [〈A0, Q0〉, · · · , 〈An, Qn〉, 〈Bi,Hi〉] is an acceptable argumentation line, n has a child ni labeled 〈Bi,Hi〉. If B = ∅ or no 〈Bi,Hi〉 ∈ B is such that Λ ′ is acceptable, then n is a leaf of the tree. The nodes of T〈A,Q〉 can be marked, yielding a tagged tree T ∗ 〈A,Q〉 as follows: • All leaves of T〈A,Q〉 are marked U in T ∗ 〈A,Q〉. • If 〈B,H〉 is the label of a node which is not a leaf, the node will be marked U in T ∗ 〈A,Q〉 if every child is marked D. Otherwise, if at least one of its children is marked U , it is marked as D. Then, given an argument 〈A, Q〉 and its associated tagged tree T ∗ 〈A,Q〉, if the root is marked U , the literal Q is said to be warranted. A is said to be the warrant for Q. Therefore, given a query Q the possible answers will be: • YES, if Q is warranted. • NO, if ¬Q is warranted. • UNDECIDED, if neither Q nor ¬Q are warranted. • UNKNOWN, if Q is not in the language of the program. Decision-Theoretic Defeat As the quick overview of DeLP shows, the key for the warrant procedure is the characterization of the defeat relation among argument structures. As we have said, specificity is introduced in the standard characterization of DeLP as an example of comparison criterion among arguments. We claim that an alternative comparison criterion may arise from decision-theoretic considerations. We will extend DeLP to allow utility values both for facts and rules. In this sense, we speak of decisiontheoretic defeasible logic programs as P ′ = (Π,∆,Φ,B) where Π and ∆ are as before, while Φ : Π ∪ ∆ → B, where B is an arbitrary Boolean algebra with top > and bottom ⊥. The new elements Φ(·) and B represent the explicit preferences of the user, in the sense that given two pieces of information μ1, μ2 ∈ Π ∪∆ if μ1 is strictly more preferred than μ2 then Φ(μ1) ÂB Φ(μ2), where oB is the order of B. The elements μ of Π ∪∆ which are most preferred receive a label Φ(μ) = >. We do not assume here that Φ assigns > to all strict rules in Π, and not even that Φ(μ1) ÂB Φ(μ2) for μ1 ∈ Π and μ2 ∈ ∆. This is because Φ(·) has, unlike the distinction between strict and defeasible rules, no epistemic content. Instead, the preferences represent other kinds of rationales, like the reliability of the source of information that provided the rule or fact, or the cost-benefit rates of the pieces of information (since their use may preclude the use of other pieces in the reasoning process), or just the inclination towards the use of certain information over another. Examples of these attitudes are pervasive even in scientific reasoning, and we will not go further into this. Of course, nothing prevents a user from giving the highest preference to strict rules and facts. Whatever the reasons are for preferring elements of Π ∪∆, the user has also to define the Boolean algebra B over which Φ(·) ranges. It can be argued that a more general ordering could be appropriate but, as we will see, the inference engine has to perform some operations over the labels of the pieces of information used in the process of argumentation. Therefore, the range of Φ(·) has to be not only an ordered set but also an algebra. In the simplest case, in which B is just a compact subset of real numbers with the natural order, we may say that Φ(μ) is the utility of the piece of information μ. From the preferences over Π ∪ ∆, we can find preferential values over defeasible derivations. Given a defeasible derivation of L from Π ∪ ∆, L1, . . . , Ln = L, let D be the set {L1, . . . , Ln} and {μ1, . . . , μn} a set of rules such that μi is the rule that yields Li. Then, that derivation yields for its conclusion L a value V (L,D) = ∧n i=1 V (Li, μi), where V (L, μ) is defined inductively as V (L, μ) = Φ(μ) if L is a fact, i.e. the body of the rule μ is empty, and V (L, μ) = Φ(μ) ∧ ∧m k=1 V (Bk, μk) if μ is a rule (strict or defeasible) with head L and body B1, . . . , Bm which is used to derive L and μk is a rule used to derive Bk. The intuition here is that a conclusion is as strongly preferred as the weakest of either its premises or the rule used in the derivation. By extension, an argument structure 〈A, L〉 yields a value for L, V (L,A) = ∧ D V (L,D), where D is a derivation that uses all the defeasible rules in A and only those defeasible rules. That is, it yields the lowest value among all the derivations of L by using defeasible rules in A. Notice that, by definition of A there is no