Weakening Commensurability Hypothesis in Possibilistic Qualitative Decision Theory
Proceedings of the Seventeenth International Joint Conference on Artificial Intelligence. Pages 717-721.
© 2001 International Joint Conferences on Artificial Intelligence
Abstract
Models for Qualitative Decision under Uncertainty assuming that uncertainty is of possibilistic nature have been proposed in 1990’s. In them, as in the classical approach of Expected Utility Theory, two alternatives emerge: `a la Von Neumann and Morgenstern (VNM) or `a la Savage. In the first one, uncertainty and preferences on consequences are measured on finite scales. Decisions are ranked in terms of the ranking induced, on their associated possibility distributions on consequences, by qualitative utility functions. To define these utilities, an hypothesis of commensurability, i.e. the existence of an onto order-preserving mapping h linking both scales of uncertainty and preference, is assumed. This hypothesis forces us to restrict to decision problems where the cardinality of the uncertainty values set is greater than or equal to the cardinality of the preference set. This point has been attacked in the possibilistic `a la Savage approach, but as far as we know, it is an open question in the qualitative VNM approach. As a first step to weaken the commensurability hypothesis in this model, the preference orderings resulting of applying the qualitative criteria without requiring h to be onto are characterized. These orderings may be not continuous, but they satisfy a relaxed continuity axiom.
- [Allais, 1953] M. Allais. Le comportement de l'homme rationnel devant le risque: critique des postulats et axiomes de l' école américaine. Econometrica,(21):503-546,1953.
- [Brafman and Tennenholtz, 1996] R.I. Brafman and M. Tennenholtz. On the foundations of qualitative decision criteria. In 13th Nat. Conf. on A.I.(AAAI'96), 1996.
- [Brafman and Tennenholtz, 1997] R.I. Brafman and M. Tennenholtz.On the axiomatization of qualitative decision criteria. In 14th Nat. Conf. on A.I.(AAAI'97), 76-81, 1997.
- [Doyle and Thomason, 1999] J. Doyle and R. Thomason. Background to qualitative decision theory. AI Magazine, 20(2):55-68, 1999.
- [Dubois and Prade, 1995] D. Dubois and H. Prade. Possibility theory as a basis for QDT. In 14th Int. Joint Conf. on A.I. (IJCAI'95), 1924-1930, 1995.
- [Dubois et al., 1997] D. Dubois, H. Prade, and R. Sabbadin. Decision under qualitative uncertainty with Sugeno integrals: an axiomatic approach. In 7th Int. Fuzzy Syst. Assoc. Cong. (IFSA'97), vol. I, 441-446, 1997.
- [Dubois et al., 1998] D. Dubois, H. Prade, and R. Sabbadin. QDT with Sugeno integrals. In 14th Conf. on Uncertainty in A.I. (UAI'98), 121-128, 1998.
- [Dubois et al., 1999] D. Dubois, L.Godo, H. Prade, and A. Zapico. On the possibilistic-based decision model:From decision under uncertainty to case-based decision. Int. J. of Uncert., Fuzz. and Know-Based Syst., 7(6): 631-670, 1999. Short version in:Making decision in a qualitative setting:from decision under uncertainty to case-based decision. In 6th Int. Conf. on Principles of Know. Repres and Reasoning (KR'98),594 -605, 1998.
- [Godo and Zapico, 2001] L.Godo and A.Zapico. On the Possibilistic-Based Decision Model: Characterisation of Preferences Relations under Partial Inconsistency. In Applied Intelligence,14(3):319-333, 2001.
- [Ellsberg, 1961] D. Ellsberg. Risk, ambiguity and the Savage axioms. Quarterly J. of Economics, 75:643-669, 1961.
- [Fargier and Perny, 1999] H. Fargier and P. Perny. Qualitative models for decision under uncertainty without the commensurability assumption. In 15th Conf. on Uncertainty in A.I. (UAI'99), 188-195, 1999.
- [Savage, 1972] L. J. Savage. The Foundations of Statistics. Dover, New York, 1972.
- [von Neumann and Morgenstern, 1944] J. von Neumann and O. Morgenstern. Theory of Games and Economic Behavior. Princeton Univ. Press, 1944.
- [Zapico, 1999] A. Zapico. Axiomatic foundations for qualitative/ordinal decisions with partial preferences. In 16th Int. Joint Conf. on A.I. (IJCAI'99), 1999. IJCAI
References were extracted automatically from the PDF and may contain errors.
Full text extracted automatically from the PDF; figures, tables and equations may differ. The PDF is the version of record.
1 Introduction
The basic references in classical Decision Theory characterizing preference relations under uncertainty and the rationality hypothesis are Von Neumann and Morgenstern's Expected Utility Theory(EUT)[1944], and the version of Savage [1972]. Both approaches assume that uncertainty is represented by probability distributions. Von Neumann and Morgenstern assume a probability distribution encoding uncertainty on situations. Then, as each decision induces a probability distribution on the set of consequences CG, each decision is identified with its associated probability distribution on CG. For sorting decisions, associated distributions are ranked in terms of their expected value with respect to Decision Maker's (DM's for short) preferences on consequences. On the other hand, Savage [1972] proposes a somewhat different framework for EUT, axiomatically characterizing the preference relation on acts of Decision Makers that behave as EUT agents.
The classical axiomatic frameworks of Utility Theory have actually been questioned rather early (e.g. [Allais, 1953; Ellsberg, 1961]). Another problem with EUT is that it needs numerical probabilities for each state and numerical utilities for all possible consequences, but this assumption may be too strong if there is only incomplete or poor available information. As Doyle and Thomason [1999] comment in a recent paper, there are many experiences showing that usually people explain and make their decisions with partial, generic and "uncertain" information. Hence, a qualitative approach may give tools for representing this decision making behaviour. They summarize main proposals on Qualitative Decision Theory. Among them, we find those models that use Possibility Theory as uncertainty formalism in which, analogously to classical EUT, two alternatives emerge: `a la Von Neumann and Morgenstern, initiated by Dubois and Prade in [1995] and latter developed and extended in [Dubois et al., 1999; Zapico, 1999], and `a la Savage, this proposal is summarized in [Dubois et al., 1997; 1998].
Here, we are interested in the possibilistic qualitative counterpart to VNM's EUT. In this approach, both DM's preferences on consequences and uncertainty are graded on finite ordinal sets (i.e. scales equipped with the maximum, minimum and an order-reversing operations) that are commensurate. As each decision is linked with its associated possibility distribution on consequences, then for ranking decisions the associated distributions are ranked by qualitative utility functions -expressed in terms of maximum, minimum and reversing operations. To define these utilities an hypothesis of commensurability, i.e. the existence of an onto order-preserving mapping linking both scales of uncertainty and preference, is assumed. This forces us to restrict to decision problems where the cardinality of the uncertainty set is greater than or equal to the cardinality of the preference set, unfortunately in several problems this cardinality relation is not satisfied.
The commensurability hypothesis has been a point of interest for researchers in the possibilistic `a la Savage approach (e.g. [Fargier and Perny, 1999]) but as far as we know, it has not been attacked in the qualitative Von Neumann and Morgenstern style. Here, we propose to weaken the commensurability hypothesis for this approach by not requiring the linking mapping scales involved to be onto. That is, we will characterize the preference ordering resulting of applying these qualitative criteria only requiring CW to be an order-preserving mapping, in particular we show that not requiring CW to be onto results in a relaxation of the continuity property of the preference relations on distributions. On the other way, replacing in the axiomatic framework the usual continuity axiom by the relaxed one is not enough for guaranteeing the relation is representable by the qualitative criteria, an axiom establishing how some special mixtures are handled is required.
This weakening of the commensurability hypothesis will allow us to deal with other types of problems, in particular, those in which the cardinality of the preference values set may be greater than the cardinality of the uncertainty set.
In the next Section we summarize the Possibilistic Qualitative Decision Theory (PQDT). Our proposal on weakening the commensurability hypothesis is introduced in Section 3, including the representation theorem for characterizing the preference relations representable by these "more general" utility functions. The respective extension of the model for involving non-normalized distributions are considered in it as well. Finally, in last Section we include some concluding remarks.
2 Background on PQDT
Let us define the PQDT framework. CE denotes a finite linear plausibility scale, where CXD2CUB4CE B5 BP BC and D7D9D4B4CE B5 BP BDBN CG is a finite set of consequences, and A5B4CGBN CE B5 denotes the set of consistent possibility distributions on CG over CE , i.e.
For the sake of simplicity, we shall use BT for denoting both a subset BT AI CG and the normalized possibility distribution on CG such that APB4DCB5 BP BD if DC BE BT and APB4DCB5 BP BC otherwise. Hence, we can consider CG as included in A5B4CGBN CE B5.
The so-called Possibilistic mixture (possibilistic lottery), the qualitative counterpart of the probabilistic lottery, is an operation defined on A5B4CGBN CE B5 that combines two possibility distributions AP BD and AP BE into a new one, denoted (ALBPAP BD BN AMBPAP BE ), with ALBN AM BE CE and D1CPDCB4ALBN AMB5 BP BDBN defined as:
With this definition, and taking into account the properties of maximum and minimum, the "reduction of lotteries" holds, i.e. B4ALBPAP BD BN AMBPB4ABBPAP BD BN ACBPAP BE B5B5 AO B4D1CPDCB4ALBND1CXD2B4AMBN ABB5B5BPAP BD BN D1CXD2B4AMBN ACB5BPAP BE B5BM
CD is a finite linearly ordered scale of preference (or utility), with D7D9D4B4CDB5 BP BD and CXD2CUB4CDB5 BP BCBN D2 CD is the reversing involution in CD. As usual, we assume as working hypothesis the existence of a preference function representing DM's preference on consequences, i.e. there exists a function D9:CG AXCD that assigns a preference level of CD to each consequence of CG such that D9B4DCB5 AK D9B4DDB5 if and only if DD is at least as preferred as DCBM Let CW:CE AX CD be an order preserving function relating CE and CD such that CWB4BCB5 BP BCBN CWB4BDB5 BP BDBM
In such a framework, assuming also that CW is onto, Dubois et al. [1999] have characterized the preference relations induced by the utility functions
where D2 BP D2 CD Æ CW. This criterion requires AP to make, at least to some extent, all plausible consequences to be preferred ones 1 . Since in some problems this criterion may be too conservative, it may be interesting to consider an optimistic behaviour, like requiring AP to make at least one of the good consequences highly plausible- at least to some extent. With this goal, Dubois et al. [1999] considered the utility function:
which is dual to CD A0 . Notice that CDB4BMCYD9B5BN restricted to CGBN coincides with D9BM
Dubois et al. have shown that 2 CD A0 has an interesting property: CD A0 preserves the possibilistic mixture, in the sense that it holds that:
They also provide the following set of axioms BTCG for preference relations on A5B4CGBN CE B5, with the max-min possibilistic mixture as the internal operation on A5B4CGBN CE B5, to characterize the ordering induced by CD A0 (BG CD A0 ).
- AF BTBD BM DA is a total pre-order(reflexive, transitive and total).
This axiom guarantees the existence of maximal and of minimal elements on B4CGBNDAB5. We denote by DC and DC a maximal and a minimal ones respectively. For each AL BE CE , let AP A0 AL BP B4BDBPDCBN ALBPDCB5 and AP B7 AL BP B4ALBPDCBN BDBPDCB5.
- AF BTBE B4D9D2CRCTD6D8CPCXD2D8DD CPDACTD6D7CXD3D2B5 BM if AP AK AP BC B5 AP BC DA APBM
- AF BTBFB4CXD2CSCTD4CTD2CSCTD2CRCTB5 BM AP BD AO AP BE B5 B4ALBPAP BD BN AMBPAPB5 AO B4ALBPAP BE BN AMBPAPB5BM
- AF BTBGB4CRD3D2D8CXD2D9CXD8DDB5 BM BKAP BE A5B4CGBN CE B5BN BLAL BE CE s.t. AP AO AP A0 AL .
BTBG establishes that distributions are preferentially equivalent to having a AL-level of uncertainty with respect to DCBM
Theorem 2.1 A preference relation B4A5B4CGBN CE B5BNDAB5 satisfies axioms BTCG if, and only if, there exist
- AF a finite linearly ordered utility scale CD, with CXD2CUB4CDB5 BP BC and D7D9D4B4CDB5 BP BD,
- AF a preference function D9 BM CG AX CD s.t. D9 A0BD B4BDB5 BIBP BN BIBP D9 A0BD B4BCB5BN
- AF an onto order-preserving function CW BM CE AX CD,
1 The utility of AP is measured in terms of a degree of inclusion of the fuzzy set AP of possible consequences into the fuzzy set of good results D9BM
2 For the sake of a simpler notation, we write CD A0 B4APB5 instead of CD A0 B4APCYD9B5 when the mapping D9 is not relevant for the context. In fact, these utility functions also depend on CW. To simplify, we omit CW and will use the notation of CD to refer a utility involving an onto CW and CD CF for the case of not requiring CW this onto condition.
in such a way that AP BC DA AP iff AP BC BG CD A0 APBN
For characterizing the optimistic behaviour representable by CD B7 , it is necessary to change the uncertainty aversion axiom BTBE by an uncertainty-prone postulate:
- AF BTBE B7 : if AP AK AP BC then AP DA AP BC BN
and to adequate the continuity axiom BTBG replacing it with:
- AF BTBG B7 BM BKAP BE A5B4CGBN CE B5BN BLAL BE CE s.t. AP AO AP B7 AL BM
In the next Section, the proposal for weakening the commensurability hypothesis is introduced.
3 A First Approach with aWeaker Commensurability Hypothesis
Let us remark that the difference with the works developed previously in [Dubois et al., 1999; 1997] is that now CW is not required to be onto. That is, given CW BM CE AX CD, an orderpreserving mapping such that CWB4BCB5 BP BCBN CWB4BDB5 BP BD, for any AP BE A5B4CGBN CE B5BN consider the qualitative utility functions
where D2 BP D2 CD Æ CWBM Observe that CD A0 CF B4A1CYD9B5 and CD A0 CF B4A1CYD9B5BN restricted to CGBN coincide with the preference function D9BM It is interesting to notice the following properties.
Lemma 3.1 These functions still preserve the mixture:
In particular, we have:
Corollary 3.2 It remains true that CD CF B4ALBPAP BD BN AMBPAP BE B5 is the median of three terms including CD CF B4AP BD B5BN CD CF B4AP BE B5BNi.e.
3.1 The Roles of the Onto Condition and of the Continuity Axiom
Let us observe that the onto condition of CW is a basic requirement involved in the characterization of the relations, since the fact of allowing CW to be a non-onto mapping results in that the continuity axiom BTBG may not be true.
Example: For instances, consider consequences are either totally possible or impossible, i.e. CE BP CUBCBN BDCVBN while preferences are measured on CD BP CUBC BO DB BO BDCV, being the set of consequences CG BP CUDCBN DC BD BN DCCVBN with DM's preference on consequences D9B4DCB5 BP BCBN D9B4DC BD B5 BP DBBN D9B4DCB5 BP BD. Since CWB4BCB5 BP BC and CWB4BDB5 BP BD, it is obvious that CD A0 CF B4APB5 BP D1CXD2 DCBECGCYAPB4DCB5BPBD D9B4DCB5BM That is, the ordering induced by CD A0 CF coincides with the maximin criterion while the ordering induced by CD B7 CF coincides with the maximax one. Observe that for AP BP CUDC BD CVBN there is no AL BE CE such that AP is preferentially equivalent to AP A0 AL BM
Hence, it is clear that the continuity property BTBG for relations on A5B4CGBN CE B5 may be lost if CW is non-onto. For facing this point, a relaxed continuity axiom is proposed. The underlying idea is to relax the continuity of the preference in the sense that there exists a subset on CG such that either the distributions are preferentially equivalent to individual consequences on this subset, or, the distributions are preferentially equivalent to having a AL-level of uncertainty with respect to DC as usual.
On the other way, having a relation B4A5B4CGBN CE B5BNDAB5 satisfying axioms BTBDBN BTBEBN BTBF and the relaxed continuity one (see BTBGCABV in Section 3.2) is not enough for guaranteeing DA is representable by CD A0 CF as it is shown following.
Example: Consider the following framework:
CG BP CUDC BS DC BS DCCVBN CG BP CUDCCVBN CE BP CUBC BO AC BO BDCVBN and consider the (reflexive and transitive) relation DA s.t. DC BS DC BS AP A0 AC BS DCBN also satisfying
All other distributions are taken equivalent to DCBM This relation satisfies axioms BTBDBNBTBEBNBTBF and BTBGCABV. But it is not representable by CD A0 CF , since although B4BDBPDCBN ACBPDCB5 BS AP A0 AC , we have CD A0 CF B4AP A0 AC B5 AK CD A0 CF B4BDBPDCBN ACBPDCB5.
3.2 Axiomatic Setting Proposed
The above discussion has led us to propose this new set of axioms BTCG for preference relations on A5B4CGBN CE B5, with the max-min mixture as the internal operation:
- AF BTBDB4D7D8D6D9CRD8D9D6CTB5 BM DA is a total pre-order .
- AF BTBEB4D9D2CRCTD6D8CPCXD2D8DD CPDACTD6D7CXD3D2B5: if AP AK AP BC B5 AP BC DA APBM
- AF BTBFB4CXD2CSCTD4CTD2CSCTD2CRCTB5 BM AP BD AO AP BE B5 B4ALBPAP BD BN AMBPAPB5 AO B4ALBPAP BE BN AMBPAPB5BM
- AF BTBGCABVB4D6CTD0CPDCCTCS CRD3D2D8CXD2D9CXD8DDB5: There exists a subset 3 CG AI CG s.t. all maximal and all minimal elements of B4CGBNDAB5 are in the complement of CG , and B4BKAP BE A5B4CGBN CE B5B5 either
- AF BTDCCXDC:
- if DCBN DD BE CG BN AC BE CE then
3 Observe that CG BP BN is possible, being BTBG recovered.
- if DC BE CG then
This axiom establishes how the relation handles (reduces) lotteries involving consequences that are not preferentially equivalent to binary lotteries of DC and DC.
Now, we introduce some results that will be useful latter.
Lemma 3.3 Axioms A1, A2, A3 , A4RC and AxMix imply BTDCBE: If BT AI CG then there is DC BE BT s.t. DC AO BTBM
Dubois et al. [1999, Lemma 1] have shown:
Lemma 1 If DA verifies axioms BTBDBN BTDCBE and BTBEBN then:
- AF DC AO B4BDBPDCBN BDBPDCB5 AO CGBM
- AF DC and DC are also the minimal and maximal elements of B4A5B4CGBN CE B5BN DAB5BM
3.3 Representation of Pessimistic Qualitative Utilities
Next, we show that the preference ordering on A5B4CGBN CE B5 induced by the qualitative/ordinal pessimistic utility CD A0 CF satisfies the above set of axioms.
Lemma 3.4
Let BG CD A0 CF be the preference ordering on A5B4CGBN CE B5 induced by CD A0 CF BN i.e., AP BG CD A0 CF AP BC iff CD A0 CF B4APB5 AK CD A0 CF B4AP BC B5BM Then BG CD A0 CF satisfies axioms set BTCGBM
Proof: BTBD is easily verified, BTBE is a consequence of maximum and minimum being non-decreasing functions.BTBF results from the fact that CD A0 CF preserves max-min possibilistic mixtures, while BTDCCXDC is not difficult to verify taking into account Lemma 3.1. Now, we check axiom BTBGCABV. If CW is onto, CG BP BN and BTBGCABV reduces to BTBGBN hence, we are in the case detailed in Section 2. If CW is non-onto, let CG BP B4CUDCCY D9B4DCB5 BE D2B4CE B5CVB5 CR BM As D9 A0BD B4BDB5 BIBP BN BIBP D9 A0BD B4BCB5, and CWB4BCB5 BP BC and CWB4BDB5 BP BDBN if DC is a maximal or a minimal element of B4CGBNBG CD A0 CF B5BN then DC BPBE CG BM We have to prove that BKAP BE A5B4CGBN CE B5 either B4BLAL s.t. CD A0 CF B4APB5 BP CD A0 CF B4AP A0 AL B5B5 or B4BLDC BE CG s.t. CD A0 CF B4APB5 BP CD A0 CF B4DCB5B5BM By definition of CD A0 CF BN BKAPBN BLDC BC BE CG, s.t. CD A0 CF B4APB5 BP D1CPDCB4D2B4APB4DC BC B5B5BN D9B4DC BC B5B5BM Hence,
- AF if CD A0 CF B4APB5 BP D2B4APB4DC BC B5B5BN then CD A0 CF B4APB5 BP CD A0 CF B4AP A0 AL B5 with AL BP APB4DC BC B5 .
- AF Otherwise, CD A0 CF B4APB5 BP D9B4DC BC B5BM Now, either D9B4DC BC B5 BE D2B4CE B5 or not. In the first option, BLAL BE CE s.t. CD A0 CF B4APB5 BP D2B4ALB5 BP CD A0 CF B4AP A0 AL B5BM While in the second one, D9B4DC BC B5 BE CG BN and CD A0 CF B4APB5 BP CD A0 CF B4DC BC B5BM A3
Now, it is shown in the following Representation Theorem that the preference orderings satisfying BTCG can always be represented by a pessimistic qualitative utility CD A0 CF BM
Theorem 3.5 A preference relation DA on A5B4CGBN CE B5 satisfies axiom set BTCG if, and only if, there exist
- (i) a finite linearly ordered utility scale CD with CXD2CUB4CDB5 BP BC and D7D9D4B4CDB5 BP BDBN
- (iii) an order-preserving 4 function CW:CE AX CD s.t. CWB4BCB5 BP BC and CWB4BDB5 BP BDBN
in such a way that AP BC DA AP iff AP BC BG CD A0 CF APBM
Proof: The "if" part corresponds to the preceding Lemma. As for the "only if" part, we go on structuring the proof, in the following three steps:
- The utility scale CD and an order-preserving function CW fromCE to CD are defined.
- A function CD A0 CF :A5B4CGBN CE B5 AXCD representingDABN i.e. s.t. CD A0 CF B4APB5 AK CD A0 CF B4AP BC B5 iff AP DA AP BC is introduced.
- The preference function D9:CG AX CD is the restriction of CD A0 CF to CG and D2 BP D2 CD Æ CW, D2 CD being the reversing involution on CDBM It is proved that
Now, we develop these steps.
- As usual, DA stratifies A5B4CGBN CE B5 in a linearly ordered set of classes of equivalently preferred distributions (AP BC BE CJAPCL iff AP AO AP BC B5BM We define CD BP A5B4CGBN CE B5BP AO with the natural (linear) order CJAPCL AK CJAP BC CL iff AP DA AP BC BM Denote by 1 and 0 the maximum and minimum elements of CD, CJDCCL BP BD and CJDCCL BP BCBM
Define the order reversing function D2:CE AX CD as D2B4ALB5 BP CJAP A0 AL CLBM BTBE guarantees that D2 reverses the order. Let D2 CD be the reversing involution in CDBN the linking-scales mapping CW is defined as CW BP D2 CD Æ D2BM
- Now, we define CD A0 CF in three steps.
- (a) For all AL BE CEBN CD A0 CF B4AP A0 AL B5 BP D2B4ALB5BM
- (b) Secondly, BKDC BE CG BN let CD A0 CF B4DCB5 BP CJDCCLBM It is easy to check that AP A0 AL DA AP A0 AL BC iff CD A0 CF B4AP A0 AL B5 AK CD A0 CF B4AP A0 AL BC B5BN and analogously, that restricted to distributions of type DCBN CD A0 CF represents DA BM
- (c) Finally, since BTBGCABV guarantees that BKAPBN either B4BLAL s.t. AP AO AP A0 AL B5 or B4BLDC BE CG s.t. AP AO DCB5BN we define
Notice that, because of BTBGCABV, CD A0 CF is well defined.
- Now, D9:CG AX CD is defined as D9B4DCB5 BP CD A0 CF B4DCB5BM Notice that D9B4DCB5 BP BD and D9B4DCB5 BP BCBM It remains to prove that CD A0 CF B4AP) = D1CXD2 DCBECG D1CPDCB4D2B4APB4DCB5B5BN D9B4DCB5B5BM With this goal, we will prove the following equalities:
4 Note that CW is not required to be onto.
By BTBGCABVBN there are several alternatives for AP BD BN AP BE BM
- (a) BLAMBN AL s.t. AP BD AO B4BDBPDCBN ALBPDCB5 and AP BE AO AP A0 AM BM
- (b) BLDCBN DD BE CG s.t. AP BD AO DC and AP BE AO DDBN
- (c) BLAL BE CEBN DC BE CG s.t. AP BD AO DC and AP BE AO AP A0 AL BN
It is not difficult to verify that applying BTBFBN BTDCCXDC and reducing lotteries, in each case we obtain that
In particular, we have that
This may be easy generalized to
- AF Now, we verify
As AP is normalized, BLDC CY BE CG s.t. APB4DC CY B5 BP BDBM Without loss of generality we assume CY BP BDBM Then, let AP CX BP B4BDBPDC BD BN APB4DC CX B5BPDC CX B5BM Since AP BP D1CPDC CXBPBDBNBMBMBMBND4 AP CX BN we have:
This ends the proof of the theorem.
A3
3.4 Representation of Optimistic Qualitative Utilities
For modeling an optimistic DM's behaviour, we consider the axioms set BTCG B7 = CUBTBDBNBTBE B7 ,BTBFBNBTBGCABV B7 BN BTDCCXDC B7 CVBN as usual being AP B7 AL BP B4ALBPDCBN BDBPDCB5, with
- AF BTBE B7 : if AP AK AP BC then AP DA AP BC BN
- AF BTBGCABV B7 : There exists a subset 6 CG AI CGBN s.t. all maximal and all minimal elements of B4CGBN DAB5 are in its complement, and BKAP BE A5B4CGBN CE B5 either B4BLAL BE CE s.t. AP AO AP B7 AL B5 or B4BLDC BE CG s.t. AP AO DCB5BM
- AF BTDCCXDC B7 :
- if DCBN DD BE CG BN AC BE CE then
- if DC BE CG then
5 Note that APB4DC BD B5 BP BDBN so D9B4DC BD B5 BP D1CPDCB4D9B4DC BD B5BN D2B4APB4DC BD B5B5BM
6 If CG BP BN, then axiom BTBG B7 is recovered.
We have analogous results to Lemmas 3.3 and 3.4, but for spaces reasons they are omitted here. The respective Representation Theorem of Optimistic Utility is:
Theorem 3.6 A preference relation DA on A5B4CGBN CE B5 satisfies axioms set BTCG B7 if, and only if, there exist
- (i) a finite linearly ordered utility scale CD with CXD2CUB4CDB5 BP BC and D7D9D4B4CDB5 BP BDBN
- (ii) a preference function D9:CG AX CD s.t. D9 A0BD B4BDB5 BIBP BN BIBP D9 A0BD B4BCB5BN
- (iii) an order preserving function CW:CE AX CD such that CWB4BCB5 BP BC and CWB4BDB5 BP BDBN
in such a way that AP BC DA AP iff AP BC BG CD B7 CF APBM
3.5 A Particular Case: Maximin andMaximax Decision Criteria
Let us consider again the simplest scale of uncertainty, CE BP CUBCBN BDCVBN that is, consequences can be either fully possible or fully impossible. This is a very particular case since for any preference scale CDBN the only requirement to be fulfilled by a mapping CW:CE AX CD is that CWB4BCB5 BP BC and CWB4BDB5 BP BDBM In this framework A5B4CGBN CE B5 is just the power set BE CG and the resulting utility functionals are
leading to the well-known maximin and maximax decision models. Now, it is very easy to check that, in order to fully characterize a preference relation on BE CG induced by these CD A0 CF and CD B7 CF BN the above axioms simplify to these ones:
- AF BTBD: DA is a total pre-order,
- AF BTBE: if BT AI BU then BU DA BTBN
- AF BTBF: if BT AO BU then BTCJBV AOBU CJBVBN
- AF BTBGCABV: BKBT AI CGBN BLDC BE CG s.t. BT AO DCBN
- AF BTDCCXDC: if DC DA DD then CUDCBN DDCV AO DCBM
Actually, in this setting, axiom BTBE is redundant since it is a logical consequence of the remaining axioms. Moreover, as we are working with a finite set CGBN BTBGCABV is a consequence of BTDCCXDCBM The axiomatic frameworks `a la Savage of these maximax and maximin criteria are provided in [Brafman and Tennenholtz, 1996; 1997].
Dubois et al. [1999] shown the necessity of extending the model to non-normalized possibility distributions, because of case-based decision-making problems or to be able of facing problems where different, partially inconsistent, information sources about the situation are available,etc. To include these problems, they extended the model summarized in Section 2. Analogously, we extend our above proposal and provide corresponding characterizations of the orderings induced by suitably modified utility functions.
3.6 Utilities for Non-Normalized Distributions
Now, we consider the set of non-necessarily normalized distributions on CG, A5 CTDC (CGBN CE ), with values on the finite uncertainty scale CE (keeping the usual definition of possibilistic mixture) as the working set of possibilistic lotteries. The utility functionals are extended to evaluate nonnormalized distributions as well. Indeed, given a mapping BY:CE AX CE s.t. BYB4BDB5 BP BCBN BKAP BE A5 CTDC (CGBN CE ):
where B4APB5 is the height of the distribution AP, B4APB5 BP D1CPDC DCBECG APB4DCB5BN and B4APB5 is a normalized version of AP defined as BDBN if APB4DCB5 BP B4APB5, and as APB4DCB5, otherwise. Notice that when B4APB5 BP BDBN the original expression is retrieved.
To characterize the preference orderings induced by CD CF in A5 CTDC (CGBN CE ) we extend BTCG and BTCG B7 respectively, with the usual [Dubois et al., 1999; Godo and Zapico, 2001] additional axiom:
- AF BTBJBY : BKAP BE A5 CTDC B4CGBN CE B5BN AP AO B4BDBPB4APB5BN BYB4 B4APB5B5BPCGB5BM Obtaining the following representation theorems.
Theorem 3.7 A preference relation DA on A5 CTDC (CGBN CE ) satisfies BTCG CTDC BP BTCG B7 BTBJBYB4resp. BTCG B7CTDC BP BTCG B7 B7 BTBJBYB5 if, and only if, there exist
- (i) a linearly ordered and finite preference scale CD with CXD2CUB4CDB5 BP BC and D7D9D4B4CDB5 BP BDBN
(ii) a preference function D9:CG AX CD s.t. u A0BD (1) BIBP BN BIBP u A0BD (0), (iii) an order-preserving mapping CW:CE AX CDBN CWB4BCB5 BP BC and CWB4BDB5 BP BDBN
4 Conclusions and Future Works
As it has been mentioned, to weaken the onto condition of the linking-scales mapping allows us to work with decisionmaking problems where the cardinality of the uncertainty values set may be not greater than or equal to the cardinality of the preference set.
It has been shown that this onto condition in the commensurability hypothesis is directly related with the continuity property of the preference relation on A5B4CGBN CE B5. For characterizing those preference relations representable by the qualitative utility functions where CW is not required to be onto, we have proposed to include a relaxed continuity axiom and an axiom for handling some special mixtures.
It is true, that given CD and CW non-onto, we may extend CE to some CE BC and define CW BC BM CE BC AX CD, with CW BC onto, in a such a way that BKAP BE A5B4CGBN CE B5BN CD CF B4APCYCWB5 BP CDB4APCYCW BC B5. But the point, for us, is that while the ordering induced by CDB4A1CYCW BC B5 in A5B4CGBN CE BC B5 satisfies continuity, the ordering induced by CD CF B4A1CYCWB5 in the original distributions set A5B4CGBN CE B5 may not satisfy continuity.
This is a first approach to the goal of weakening commensurability. A next step is to extend the analysis to the general utility functions proposed by Dubois et al. [1999]. However, this extension seems non-trivial since the onto conditions is required to guarantee the well-definedness of these generalized utility functions. Another point is to consider the utilities when the sets values of uncertainty and preferences are non-linear lattices [Zapico, 1999].
Acknowledgments
The author wishes to thank the valuable comments of Llu´ıs Godo and of the anonymous referees .
@inproceedings{ijcai2001p98,
title = {Weakening Commensurability Hypothesis in Possibilistic Qualitative Decision Theory},
author = {Zapico, Adriana},
booktitle = {Proceedings of the Seventeenth International Joint Conference on
Artificial Intelligence, {IJCAI-01}},
publisher = {International Joint Conferences on Artificial Intelligence},
editor = {Bernhard Nebel},
pages = {717--721},
year = {2001},
url = {https://ijcai.org/proceedings/2001/98},
}
TY - CPAPER TI - Weakening Commensurability Hypothesis in Possibilistic Qualitative Decision Theory AU - Zapico, Adriana T2 - Proceedings of the Seventeenth International Joint Conference on Artificial Intelligence, IJCAI-01 PB - International Joint Conferences on Artificial Intelligence PY - 2001 SP - 717 EP - 721 UR - https://ijcai.org/proceedings/2001/98 ER -Copied
Machine-readable versions of this paper, for tools and AI agents: