Elsevier

Fuzzy Sets and Systems

Volume 304, 1 December 2016, Pages 45-64
Fuzzy Sets and Systems

When upper conditional probabilities are conditional possibility measures

https://doi.org/10.1016/j.fss.2016.02.015Get rights and content

Abstract

Conditioning for (non-additive) uncertainty measures is still an open problem. This is essentially due to the fact that these measures can be viewed either as lower or upper probabilities or axiomatically defined, independently of probability theory. Focusing on possibility measures on finite domains, the aim is to generalize to the conditional case the interpretation of possibility measures as upper envelopes of extensions of suitable probability measures.

Introduction

Probability theory has a fundamental role for dealing with uncertainty, even if sometimes the decision maker has only an incomplete, partial, imprecise probabilistic information. In such situations it is unavoidable to consider a set of probabilities together with the corresponding envelopes, which turn out to be non-additive uncertainty measures (see, e.g., [12], [20], [21], [27], [43]).

For example, prior probabilities are crucial in Bayesian statistics, even if sometimes the elicitation of the prior is not an easy task. For this, in some applications (for instance, in decision problems with multiple agents [15], in the reconstruction of images or when misclassified variables are present [38]) instead of assessing a single prior probability, one can select a suitable family of priors.

Concerning the relationship between probability theory and the chosen family of non-additive measures, it is well-known that the upper [lower] envelope of the coherent extensions of a coherent probability is subadditive [superadditive] but is not necessarily 2-alternating [2-monotone] (see, for instance, [13], [43] and Example 1 in [2]). Nevertheless, starting from a probability on an algebra the upper [lower] envelope of its extensions to another algebra is a plausibility [belief] function (see [9], [18], [20], [22], [28], [43]). In [25] it is provided a characteristic property for an upper probability to coincide with a possibility measure, which led in [10], [11], [12] to characterize possibility measures as upper envelopes of the extensions of a probability when a logical condition (weak logical independence) holds among the involved algebras.

Starting from the seminal works [24], [47], possibility measures have received a lot of attention, due to their key role in fuzzy set theory, soft computing and decision theory. The debate on the conditioning for possibility measure or, more generally, for non-additive uncertainty measures is still an open issue, even if several conditioning notions have been introduced in literature (see, e.g., [1], [7], [14], [16], [20], [24], [26], [30]).

A first aspect in which the given conditioning notions differ is how the conditional measure is defined: essentially, either starting from a non-additive unconditional measure through a suitable operation or directly as a function satisfying a set of axioms. Furthermore, conditional uncertainty measures can be viewed as specific lower or upper conditional probabilities, so the relation with probability theory can be investigated. In this line, in [5] we have generalized the Bayesian rule of conditioning for belief functions on finite domains, originally proposed by [20], [28], [32].

In this paper we go further in this investigation by studying when conditional plausibility functions are conditional possibility measures. In detail, focusing on a finite Boolean algebra A, we introduce the notion of Bayesian conditional possibility (or B-conditional possibility for short) which is obtained through a generalization of the Bayesian conditioning rule [20], relying on a linearly ordered class of possibility measures. For any B-conditional possibility measure ΠB(|) defined on A×A0, with A0=A{}, there exists a finite Boolean algebra A and a conditional probability P(|) defined on A×A0 such that the upper envelope of its extensions coincides with ΠB(|) on A×A0. Furthermore, we investigate the conditions under which the upper envelope of the extensions of a full conditional probability is a full B-conditional possibility. It is shown that weak logical independence of two finite algebras A and A, is a sufficient (but not necessary) condition for the upper envelope of the conditional probabilities extending a conditional probability P(|) defined on A×A0 to be a B-conditional possibility on A×A0. Contrary to the unconditional case, there are B-conditional possibilities ΠB(|) on A×A0 for which there is no finite algebra A, weakly logically independent of A, such that a conditional probability on A×A0 gives rise to ΠB(|) as upper envelope of its extensions (see Example 6). Thus we introduce a weaker form of weak logical independence that allows us to provide a necessary and sufficient condition for the upper envelope of the extensions of a conditional probability P(|) defined on A×A0 to be a B-conditional possibility measure on A×A0.

The paper is structured as follows. In Section 2 some preliminary notions on conditional probabilities and their extensions are recalled, while in Section 3 the problem of conditioning for possibility measures is introduced. Section 4 presents the generalized Bayesian conditioning rule which gives rise to full B-conditional possibilities. Section 5 copes with the logical notion called weak logical independence. Such condition implies a conditional form of nesting giving rise to a full B-conditional possibility as the upper envelope of the extensions of any full conditional probability. Since weak logical independence turns out to be only a sufficient condition for obtaining a full B-conditional possibility, in Section 6 a necessary and sufficient condition is given. Finally, in Section 7 some conclusions are drawn.

Section snippets

Conditional probability and its enlargement

Let A be a Boolean algebra of events E's and denote by ()c, ∨ and ∧ the usual Boolean operations of contrary, disjunction and conjunction, respectively. The sure event Ω and the impossible event ∅ coincide, respectively, with the top and bottom elements of A endowed with the partial order ⊆ of implication.

A conditional event E|H is an ordered pair of events with H, where the conditional event E|Ω is customarily identified with E. Let H be an additive class (i.e., a set of events closed under

Conditioning for possibility measures

Let A be a finite Boolean algebra with set of atoms CA={D1,,Dm}. A possibility measure Π on A is a normalized maxitive function [24], i.e., it satisfies condition (N) of the previous section together with

  • (Π)

    Π(AB)=max{Π(A),Π(B)}, for every A,BA.

Every possibility measure Π has an associated dual necessity measure N, defined for every EA as N(E)=1Π(Ec), which is minitive, i.e., for every A,BA,N(AB)=min{N(A),N(B)}.

Possibility measures are particular plausibility functions [41] and so

Generalized Bayesian conditioning

The aim of this section is to generalize the Bayesian conditioning rule in equation (6) to cover the cases where the denominator vanishes, by preserving the upper envelope interpretation of equation (7). The idea is to provide a generalized Bayesian conditioning rule relying on a linearly ordered class {Π0,,Πk} of possibilities on A such that the resulting conditional possibility ΠB(|) can be obtained through the extension of a full conditional probability P(|) defined on a different

Weakly logically independent Boolean algebras

In this section we recall the condition of weak logical independence introduced in [10], [11], [12] in order to show its role in the properties of the upper envelope of the extensions of a full conditional probability.

Definition 4

A finite Boolean algebra A with set of atoms CA={D1,,Dm} is weakly logically independent of another finite Boolean algebra A with set of atoms CA={C1,,Cn} (in symbol AwA) if, for any given DjCA, every DkCA (kj) satisfies at least one of the following conditions:

  • (i)

    Dk(Dj);

A necessary and sufficient condition

In the previous section it has been shown that weak logical independence of A and A is only a sufficient condition in order to get a full B-conditional possibility ΠB(|) on A as upper envelope of the extensions of a full conditional probability P(|) on A.

In case of an unconditional possibility measure Π() on A, in [12] it has been shown that it is always possible to find a weakly logically independent finite Boolean algebra A and an unconditional probability measure P() on A giving

Conclusions

The paper proposes a notion of conditioning (B-conditioning) for possibility measures on finite algebras which generalizes the Bayesian conditioning rule. Such definition is not axiomatic since it relies on the existence of a (non-unique) linearly ordered class of possibility measures (C-class). The resulting conditional measure (full B-conditional possibility) is shown to be the upper envelope of the extensions of a full conditional probability defined on a different algebra. The purely

Acknowledgements

This work was partially supported by INdAM–GNAMPA through the Project 2015 U2015/000418Envelopes of coherent conditional probabilities and their applications to statistics and artificial intelligence”, and by the Italian Ministry of Education, University and Research, funding of Research Projects of National Interest (PRIN 2010–2011) under the grant 2010FP79LR_003Logical methods of information management”.

References (47)

  • P. Walley et al.

    Coherence of rules for defining conditional possibility

    Int. J. Approx. Reason.

    (1999)
  • P.M. Williams

    Notes on conditional previsions

    Int. J. Approx. Reason.

    (2007)
  • L.A. Zadeh

    Fuzzy sets as a basis for a theory of possibility

    Fuzzy Sets Syst.

    (1978)
  • B. Bouchon-Meunier et al.

    Independence and possibilistic conditioning

    Ann. Math. Artif. Intell.

    (2002)
  • G. Coletti et al.

    Bayesian inference: the role of coherence to deal with a prior belief function

    Stat. Methods Appl.

    (2014)
  • G. Coletti et al.

    When upper conditional probabilities are conditional possibility measures

  • G. Coletti et al.

    Characterization of coherent conditional probabilities as a tool for their assessment and extension

    Int. J. Uncertain. Fuzziness Knowl.-Based Syst.

    (1996)
  • G. Coletti et al.

    From conditional events to conditional measures: a new axiomatic approach

    Ann. Math. Artif. Intell.

    (2001)
  • G. Coletti et al.

    Probabilistic Logic in a Coherent Setting

    (2002)
  • G. Coletti et al.

    Toward a general theory of conditional beliefs

    Int. J. Intell. Syst.

    (2006)
  • G. Coletti et al.

    Possibility measures in probabilistic inference

  • G. Coletti et al.

    Possibility measures through a probabilistic inferential process

  • G. Coletti et al.

    Representability of ordinal relations on a set of conditional events

    Theory Decis.

    (2006)
  • Cited by (20)

    • The extent of partially resolving uncertainty in assessing coherent conditional plausibilities

      2023, Fuzzy Sets and Systems
      Citation Excerpt :

      Hence, the notion of probabilistic coherence acts like a “bridge” between finitely additive probability theory and other uncertainty calculi. For instance, in some particular cases, due to specific logical conditions, these envelopes can be belief and plausibility functions [8,16,36] or even necessity and possibility measures [9,26]. The paradigm of coherence has been extended referring to a class of more general uncertainty measures giving up on additivity (see, e.g., [63,61,62,64,65,15,18,4,29,43]).

    • The price of risk based on multilinear measures

      2022, International Review of Economics and Finance
    • Modeling agent's conditional preferences under objective ambiguity in Dempster-Shafer theory

      2020, International Journal of Approximate Reasoning
      Citation Excerpt :

      Generally, the lower envelope of a class of probability measures is only a superadditive capacity. Nevertheless, under suitable conditions the lower envelope can be a belief function or even a necessity measure [10,11,16,27,44]. However, given a lower probability it is always possible to look for a belief function outer approximating it [45].

    View all citing articles on Scopus
    View full text