Revising beliefs on the basis of evidence (Q432970): Difference between revisions

This paper addresses the problem of modelling how the belief set of an agent is modified in order to incorporate some new information, in the cases when such input is (considered) uncertain rather than categorical. More precisely, it proposes an approach which accounts for the case when the new information (and, in fact, each one of the agent's beliefs in general) is not simply a sentence (which is assumed to be certain) but rather a sentence and the probability that is associated to it being true. In this paper a propositional language based on a finite set of atomic sentences, closed under the usual Boolean connectives, is assumed. In the approach presented, the epistemic state \(K\) of an agent is modelled as a pair \((P,c)\), where \(P\) is a probability assignment over possible worlds and \(c \in (0,1]\) is a {confidence level}. Given such an epistemic state, the agent's (categorical) accepted beliefs are defined in terms of a set of possible worlds, denoted by \(\mathrm{Mod}(K)\), which can be informally described as the least set of worlds of greatest probability such that the probability of that set is greater than or equal to \(c\). Thus, the set of the agent's (categorical) accepted beliefs, \(\mathrm{Bel}(K)\), consists of the formulas of which every world in \(\mathrm{Mod}(K)\) is a model. Based on such formal framework, the paper proposes a way of updating the agent's probability assignment to possible worlds when it receives some uncertain information, i.e., an input of the form \((\phi,q)\), where \(\phi\) is a formula and \(q \in [0,1]\) is the probability of \(\phi\) being true. More precisely, given an epistemic state \(K = (P,c)\), and an input \((\phi,q)\), a new probability assignment to possible worlds \(P(\phi,q)\) is defined from \(P\) via Bayesian conditioning. After that, the interrelation between the probabilistic framework presented and some other approaches to belief revision is analysed. In that regard, first of all the notion of revision (expressed as a binary function) of an epistemic state is introduced as follows: The (outcome of the) revision of \(K = (P,c)\) by (a formula) \(\phi\) is the epistemic state defined by \(K * \phi = (P(\phi,q),c)\), where \(q\) is assumed to be a fixed number in the range \((0.5, 1.0)\). Afterwards it is clarified that, of the AGM revision postulates (K\(*\)1)--(K\(*\)8) [\textit{C. E. Alchourrón}, \textit{P. Gärdenfors} and \textit{D. Makinson}, J. Symb. Log. 50, 510--530 (1985; Zbl 0578.03011)], such probability-based belief revision only satisfies (K\(*\)1), (a slightly modified version of) (K\(*\)5) and (K\(*\)6). In this respect it is highlighted that the postulate of success (viz. (K\(*\)2)) is not guaranteed to hold but, since if \(\phi\) is possible -- in that there is a world \(w\) such that \(w\) is a model of \(\phi\) and \(P(w) > 0\) -- an agent will accept \(\phi\) after a sufficient number of iterations (i.e. of successive revisions of its epistemic state by \(\phi\)), \(*\) is a belief revision operator. It is also emphasised that, since (K\(*\)3) and (K\(*\)4) are not guaranteed to hold, revision by a formula consistent with the agent's belief set does not necessarily correspond to expansion. Still concerning this issue, it is remarked in the paper that a small modification to the approach proposed is enough to assure that, in addition, the postulates (K\(*\)3) and (K\(*\)7) hold. Additionally, the approach presented is also analysed with respect to iterated revision and it is shown that of the postulates (C1)--(C4) of \textit{A. Darwiche} and \textit{J. Pearl} [Artif. Intell. 89, No. 1--2, 1--29 (1997; Zbl 1018.03012)] only (C3) is satisfied, and also that the independence postulate (Ind) of \textit{Y. Jin} and \textit{M. Thielscher} [Artif. Intell. 171, No. 1, 1--18 (2007; Zbl 1168.03318)] does not hold. The cause for this is exposed. Finally, it must be noted that the paper contains yet a detailed exposition of the relation among the approach presented here and some other approaches in belief revision, namely the ones presented in [\textit{S. Shapiro} et al., ``Iterated belief change in the situation calculus'', in: A. G. Cohn (ed.) et al., Proceedings of the seventh international conference on the principles of knowledge representation and reasoning. San Francisco: Morgan Kaufmann. 527--538 (2000); \textit{A. C. Nayak}, ``Iterated belief change based on epistemic entrenchment'', Erkenntnis 41, No. 3, 353--390 (1994; \url{doi:10.1007/BF01130759}); \textit{S. Konieczny} and \textit{R. P. Pérez}, ``Improvement operators'', in: G. Brewka (ed.) et al., Proceedings of the eleventh international conference on the principles of knowledge representation and reasoning, Sydney, Australia, 2008. Menlo Park, CA: AAAI Press. 177--186 (2008); \textit{P. Gärdenfors}, Knowledge in flux. Modeling the dynamics of epistemic states. Cambridge, MA: MIT Press (1988; Zbl 1229.03008), Chapter 5; \textit{S. Lindström} and \textit{W. Rabinowicz}, J. Philos. Logic 18, No. 1, 69--101 (1989; Zbl 0669.03007); \textit{C. Boutilier}, Notre Dame J. Formal Logic 36, No. 1, 158--182 (1995; Zbl 0844.03016); \textit{F. Bacchus} et al., Artif. Intell. 111, No. 1--2, 171--208 (1999; Zbl 0996.68192)].