ISSN 10645624, Doklady Mathematics, 2011, Vol. 83, No. 3, pp. 418–420. © Pleiades Publishing, Ltd., 2011. Original Russian Text © V.D. Noghin, O.V. Baskov, 2011, published in Doklady Akademii Nauk, 2011, Vol. 438, No. 4, pp. 456–459. COMPUTER SCIENCE Pareto Set Reduction Based on an Arbitrary Finite Collection of Numerical Information on the Preference Relation V. D. Noghin and O. V. Baskov Presented by Academician S.K. Korovin January 21, 2011 Received February 17, 2011 DOI: 10.1134/S1064562411030288 Many applied problems in economics and engi neering can be formulated as a multicriteria choice problem with several numerical functions. A specific feature of multicriteria problems is that, starting a choice procedure, the decision maker (DM) cannot, as a rule, precisely express his or her interests and pref erences, which underlie the choice made. Thus, beginning the search for a set (in a special case, a sin gleton) of “best” elements, the DM does not have the exact definition of this concept. Frequently, these best elements are detected in the course of decision making based on available information about the DM’s prefer ences. Numerous procedures and methods have been pro posed for solving multicriteria problems depending on the type and character of information on the DM’s preferences [1, 2]. Frequently, these are heuristic pro cedures that yield substantially different best solutions. According to the overwhelming majority of research ers, best solutions have to be sought in the set of Pareto optimal (effective, tradeoff) alternatives. This circum stance is expressed in the Edgeworth–Pareto princi ple, which has relatively recently been axiomatically substantiated [3]. Thus, the problem of choosing a set of best alternatives can be reformulated as the problem of Pareto set reduction. Consider the model of multicriteria choice [3], which contains a set X of initial alternatives, a vector criterion y = f(x) = (f1(x), f2(x), …, fm(x)), and an asymmetric binary preference relation X defined on X. Let Y = f(X), and let Y be a binary relation on the set Y induced by the relation X as follows: x1 X x2 ⇔ f(x1) Y f (x2) for all x1 ∈ x̃ 1 , x2 ∈ x̃ 2 ; x̃ 1 , x̃ 2 ∈ X˜ , where X˜ is the collection of equivalence classes gener ated by the equivalence relation x1 ~ x2 ⇔ f(x1) = f(x2) St. Petersburg State University, Universitetskii pr. 28, Petrodvorets, St. Petersburg, 198504 Russia email: [email protected] on X. The sets of chosen (best) alternatives and vectors are denoted by C(X) and C(Y) = f (C(X)), respectively. These are the sets to be determined in the course of making a choice. The set of Pareto optimal alternatives with respect to the vector criterion f on X is denoted by Pf(X). The Pareto set Pf(X) can be reduced (i.e., certain Paretooptimal elements can be eliminated) if we have additional information on the DM’s preferences. The most reliable and simple from a practical point of view is information on the DM’s readiness to make a cer tain compromise. Frequently, this compromise means that the DM agrees to lose according to insignificant criteria while gaining according to significant criteria. Taking into account information on the DM’s preference relation, which the DM is usually initially totally unaware of, underlies an axiomatic approach to the reduction of the Pareto set. This approach has been developed for almost three decades [3]. First, it was established how one should use a single piece of information (“information quantum”) in the form of two collections of numbers, of which one indicates the admissible upper limits of losses for the group of insig nificant criteria and the other shows the gain values for the significant criteria that are larger than or equal to those the DM would agree to receive when making the compromise. Later, possibilities of using various col lections of such information (several information quanta) for Pareto set reduction were studied. In the geometric language, an information quan tum about the DM’s preference relation means [3] that the relation u Y 0 holds for some u ∈ Nm, where Nm is a set of mdimensional vectors with at least one strictly positive and one strictly negative component. The specification of a collection of information quanta is equivalent to the fulfillment of the relations ui Y 0, i = 1, 2, …, k. In this paper, we propose two universal algorithms (methods) based on taking into account an arbitrary finite collection of information quanta for Pareto set reduction. The first (geometric) algorithm was devel oped by the first author. It involves the solution of the 418 PARETO SET REDUCTION BASED ON AN ARBITRARY FINITE COLLECTION convex analysis problem posed in [3]. The second (algebraic) algorithm was developed by the second author. It sequentially takes into account the available collection of quanta on the basis of the linear operator used to take into account one quantum in Theorem 3.5 from [3]. In what follows, four reasonable choice axi oms formulated in [3] are assumed to hold. The above geometric problem of convex analysis is stated as follows. Given a finite collection of vectors a1, a2, …, ak ∈ Rm, k > m, generating an acute convex mdimensional cone M, the goal is to construct a min imal (in number) collection of vectors b1, b2, …, bn that generate the dual {y ∈ Rm | 〈ai, z〉 ≥ 0, i = 1, 2, …, k} of the cone M. Here, the angle brackets denote the scalar product of vectors. It should be noted that the sought vectors b1, b2, …, bn always exist and are determined up to a positive multiplier. The following algorithm is proposed for solving this problem. The collection of vectors a1, a2, …, ak is fed as input, while the output (in storage) is the desired collection of vectors. Algorithm 1 Step 1 (start of a loop over the vectors). Start a loop m–1 in which all possible (m – 1) over i from 1 to C k vector subsets of the collection a1, a2, …, ak are gener ated. Step 2 (verification of linear independence). If the current ith subset ai1, ai2, …, ai(m – 1) chosen from a1, a2, …, ak is linearly dependent, then increase the number i by one and go to the beginning of Step 2. If i cannot m–1 be further increased, i.e., i = C k , go to Step 5. Oth erwise, i.e., if the subset is linearly independent, go to Step 3. Step 3 (construction of an orthogonal vector). Form a square matrix D of order n from the column vectors of the subset ai1, ai2, …, ai(m – 1). Preliminarily, any vector from the set Ii = {a1, a2, …, ak}\{ai1, ai2, …, ai(m – 1)} that form, together with ai1, ai2, …, ai(m – 1), a linearly independent system is written to the columns on the right. Find the last column of the matrix (DT)–1, where T denotes the transpose. The resulting column vector (denote by yi) is stored. Step 4 (verification of whether the vector y i belongs to the desired set b1, b2, …, bn). Calculate the scalar products 〈a j, yi〉 for all vectors a j ∈ Ii. If at least one scalar product is negative, remove y i from the storage. Increase i by one and go to Step 2 (if this increase is not possible, go to Step 5). Remark. To reduce the search and eliminate iden tical (up to a positive multiplier) desired vectors from the storage, for each stored vector yi, the correspond ing collection Yi of all vectors a1, a2, …, ak orthogonal to y i is stored at Step 4 (i.e., those a j for which 〈aj, yi〉 = 0). At Step 2, if the current subcollection ai1, ai2, …, ai(m – 1) is a subset of at least one previously DOKLADY MATHEMATICS Vol. 83 No. 3 2011 419 formed set Yi, skip this subcollection and increase i by one. Step 5 (termination of the algorithm). After exe cuting a complete cycle over i, the desired column vec tors b1, b2, …, bn (denoted by yi) are stored. Theorem 1. The vectors b1, b2, …, bn produced by finite Algorithm 1 form a minimal collection of vectors generating a cone that is the dual of the mdimensional acute convex cone M generated by the collection of vec tors a1, a2, …, ak ∈Rm, k > m. Theorem 2. Let four reasonable choice axioms hold and a consistent collection of information quanta ui Y 0, i = 1, 2, …, k be given [3]. Then, for any set of selected alternatives C(X), C ( X ) ⊂ P g ( X ) ⊂ P f ( X ). (1) Here, Pg(X) is the set of Pareto optimal alternatives on the original set X with respect to the new ndimensional vector criterion g(x) =(〈b1, f(x)〉, 〈b2, f(x)〉, …, 〈bn, f (x)〉), n ≥ m, where the vectors b1, b2, …, bn are produced by Algorithm 1 as applied to the collection consisting of vec tors specifying the information quanta u1, u2, …, uk together with m unit basis vectors of the space Rm. According to Theorem 2, when Algorithm 1 is applied, it is necessary to construct a new vector crite rion g with respect to which the new Pareto set gives upper bound (1) for the unknown set of selected alter natives C(X) in view of the available collection of information quanta. The algebraic approach maximally takes into account the specific features of the given collection ui Y 0, i = 1, 2, …, k, and is sequential; i.e., informa tion is taken into account strictly in the order of avail able indexed quanta. Given a vector u1 = (u1, u2, …, um), we introduce the sets A = {i| ui > 0}, B = { j | uj < 0}, and C = {l| ul = 0}. If the sets A and B are nonempty, the vector u1 generates an information quantum about the preference relation if u1 Y 0. To take it into account, Theorem 3.5 in [3] is used to construct a new vector criterion g whose components are related to those of the criterion f by the following relations (the order of the components is not important): gi = fi g ij = u i f j – u j f i ∀i ∈ A ∪ C, ∀( i, j ) ∈ A × B. In matrix form, these relations can be rewritten as g = Tf, where T is a rectangular matrix of suitable size. The criterion g thus constructed generates a new set of vectors Z = g(X), on which a new relation Z is induced that is associated with the old as follows: x' X x'' ⇔ g(x') Z g(x''). In what follows, for brevity, the index (the symbol of the set on which the consid ered relation acts) is omitted. If necessary, it can easily be recovered from the context. It can be proved that the relation induced on the new set of vectors has all 420 NOGHIN, BASKOV the properties guaranteeing the fulfillment of four rea sonable choice axioms. Now, suppose that we are given a collection of information quanta ui 0, i = 1, 2, …, k. After taking into account the first quantum (as described above) and constructing the linear mapping T, the remaining vectors ui ∈ Y in the new set of admissible vectors are assigned their images (vectors) Tui ∈ Z. It can be shown that ui 0 ⇔ Tui 0, i = 2, 3, …, k. This opens up the possibility of taking into account each subse quent information quantum in the image space of T. For Tui 0, i = 2, 3, …, k, to be treated as relations specifying information quanta, it is necessary that each of the vectors Tui have at least one positive and at least one negative component. If there is a vector with all nonpositive components, then the original collec tion of quanta is inconsistent; i.e., the relations ui 0 cannot hold simultaneously within the framework of the reasonable choice axioms. If among Tui there is a vector with nonnegative components, then it can be discarded, since it does not carry additional informa tion, because, under the indicated axioms, any non zero vector v with nonnegative components satisfies the relation v 0. Thus, the problem of consecutively taking into account s information quanta on the preference rela tion is reduced to the problem of taking into account the preceding s – 1 quanta with the subsequent appli cation of Theorem 3.5 from [3] to the corresponding images. Let us describe Algorithm 2 in consecutive form. Assume that, by the step indexed by s ≥ 1, s – 1 infor mation quanta have been taken into account, while the subsequent quanta Ts – 1…T1ui 0, i = s, s + 1, …, k remain unaccounted. The lefthand side of the last relation represents the product of the matrices Ts – 1, Ts – 2, …, T1 and the vector ui. Step s.3. To reduce the subsequent computations, discard the vectors Ts – 1…T1ui ≥ 0. Step s.4. If among the vectors constructed at Step s.2, there are ones all of whose components are nonpositive, then terminates Algorithm 2 with the message that the original collection of information quanta is inconsistent. Step s.5. If there are remaining unaccounted quanta, go to Step s + 1. Otherwise, terminate Algo rithm 2 with the output being the matrix TsTs – 1…T1. Theorem 3. Let four reasonable choice axioms be satisfied [3]. If the collection of information quanta ui Y 0, i = 1, 2, …, k, is inconsistent, then Algorithm 2 termi nates with an inconsistency message. If this collection of quanta is consistent, then Algorithm 2 terminates, pro ducing a matrix Q such that, for any set of selected alter natives C(X), inclusions (1) hold true, where Pg(X) is the set of Pareto optimal alternatives on the set X with respect to the vector criterion g(x) = Q f (x). Moreover, the set Pg(X) is independent of the indexing of information quanta. Thus, Algorithm 1 produces a minimal (in the number of components) vector criterion g for which inclusions (1) hold. Algorithm 2 generates a new vec tor criterion g that, generally speaking, differs from that produced by Algorithm 1, and the dimension of the last criterion may appear to be noticeably higher than the minimal one. We intend to continue the study of the above algorithms, specifically, to modify Algo rithm 2 so as to eliminate the “redundant” compo nents of the vector criterion in the course of its opera tion. ACKNOWLEDGMENTS This work was supported by the Russian Founda tion for Basic Research, project no. 110700449a. REFERENCES Algorithm 2 Step s.1. According to Theorem 3.5 in [3], con struct a linear transformation Ts taking into account the information quantum Ts – 1Ts – 2…T1us 0. Step s.2. Construct vectors Ts Ts – 1…T1ui, i = s + 1, …, k. 1. A. B. Petrovskii, Decision Making Theory (Akademiya, Moscow, 2009) [in Russian]. 2. V. D. Noghin, Iskusstv. Intellekt Prinyatie Reshenii, No. 1, 98–112 (2008). 3. V. D. Noghin, Decision Making in Multicriteria Envi ronment: A Quantitative Approach (Fizmatlit, Moscow, 2005) [in Russian]. DOKLADY MATHEMATICS Vol. 83 No. 3 2011
© Copyright 2025 Paperzz