Two Approaches to Belief Revision
Transcription
Two Approaches to Belief Revision
Setup EUT Revision AGM Revision Comparison Future Work References Extras Setup EUT Revision AGM Revision Comparison Future Work References Extras Today, we’ll sketch a new approach to (qualitative) belief revision based on epistemic utility theory (EUT) and contrast it with the traditional AGM theory of belief revision. Two Approaches to Belief Revision Ted Shear1 The EUT approach is based on a (normative) Lockean thesis. It’s well known (lottery and preface paradoxes, etc. [2]) that Lockean approaches to full belief fail to satisfy Cogency. Branden Fitelson2 1 Philosophy Cogency. An agent’s belief set B should (at any given time) be deductively consistent and closed under logic. @ UC Davis Our main focus will be on divergences between EUT & AGM that are orthogonal to the classic debates about Cogency. [email protected] 2 Philosophy & RuCCS @ Rutgers As we will see, there is a precise sense in which AGM is more demanding on the agent’s commitments than EUT. [email protected] The upshot of this will be that — from an EUT perspective — AGM is an epistemically risk-seeking strategy for revising one’s qualitiative beliefs. We begin with some setup. Shear & Fitelson Setup 1 Two Approaches to Belief Revision EUT Revision AGM Revision Comparison Future Work References Extras Our agents possess both numerical credence functions, b(·), and qualitative belief sets, B. When p ∈ B, we write B(p). EUT Revision AGM Revision Comparison Future Work References Extras We take a veritistic (i.e., accuracy-centered) approach to epistemic utility according to which the only feature of epistemic attitudes that matters is their accuracy. (1) b(·) is a classical (Kolmogorov) probability function. (2) given a prior b(·), the posterior b0 (·) is generated via conditionalizing b(·) on E — i.e., b0 (·) = b(· | E). More precisely, we will adopt the following naïve, accuracy-centered epistemic utility function for belief. r if p is true at w u(B(p), w) Ö −w if p is false at w On the belief side, we’ll assume our agent entertains (classical, possible world) propositions in some finite agenda A, generated by a classical sentential language. (3) B is the set of all and only those members of A which our agent believes; and The only constraint we will impose on r and w is (4) given a prior belief set B, the posterior belief set B0 is generated by revising the prior by E — i.e., B0 = B ? E (where ? is some qualitative belief revision operator). Two Approaches to Belief Revision Setup 2 Two Approaches to Belief Revision The basic idea tenet underlying EUT is that an agent’s belief set should (at any given time) maximize expected epistemic utility from the point of view of her credence function. We will be talking about agents who revise both their credence functions and their belief sets, upon learning (exactly) some proposition E. On the credence side: Shear & Fitelson Shear & Fitelson (†) 1 ≥ w > r ≥ 0. The (Lockean) rationale for (†) will be explained shortly. 3 Shear & Fitelson Two Approaches to Belief Revision 4 Setup EUT Revision AGM Revision Comparison Future Work References Extras Setup The expected epistemic utility (EEU) of a belief B(p), from the point of view of a credence function b(·), is given by X EEU (B(p), b) Ö b(w) · u(B(p), w) Setup w . r+w Two Approaches to Belief Revision EUT Revision AGM Revision Comparison Future Work References Formally, B0 = B ÷ E, where n B ÷ E Ö p b(p | E) > 5 Extras Shear & Fitelson Setup w r+w o . 6 Two Approaches to Belief Revision EUT Revision AGM Revision Comparison Future Work References Extras w w and b(q) > r+w (i.e., that Proposition. Suppose b(p) > r+w w 1 our EUT agent believes both p and q) and /2 < r+w ≤ 1. (P2) If an agent learns something that she already believes, then her belief set should remain unchanged. Then, b(p | q) > 0 [More formally, X ∈ B ⇒ B = B ? X = B.] w−r w , and b(p | q) − b(p) < r2 rw+w2 . And, in the limit as an agent’s credences in p and q approach 1, (P2) will be satisfied by EUT revision. Informally, the reason ÷ violates (P2) is that — even if an agent already believes q — learning q can lower her credence in some other p proposition she also believes. This goes some way toward explaining why (P2) may seem like a plausible diachronic constraint on full belief, since it is “approximately” true if full beliefs have sufficiently high credence (and it is exactly true in the extremal case). Indeed, learning something one already believes (q) can drop one’s credence in another proposition one also believes (p) below one’s EUT Lockean threshold. More generally, extremal EUT agents (i.e., agents such that w = 1 and r = 0, who will have Lockean thresholds equal to one) will always satisfy all AGM principles. However, there is a precise upper-bound on the “degree” to which (P2) can fail from the perspective of EUT. Two Approaches to Belief Revision Extras The following proposition provides a bound on how much an EUT agent’s credence in one of her beliefs can be lowered by learning something else that she already believes. To get a feel for how EUT Revision works, it is instructive to note that ÷ does not generally satisfy the following (AGM) principle, which has recently been employed by Leitgeb [6]. Shear & Fitelson References EUT Revision. If an agent with a prior belief set B learns (exactly) E, then her posterior B0 should maximize EEU relative to her conditional credence function b(· | E). Theorem (Dorst [3], Easwaran [4]) A belief set B (on A) maximizes EEU relative to b if and only if, for every p ∈ B Shear & Fitelson Future Work Our diachronic requirement will be that — upon learning E via conditionaliation — our agent beleives exactly those propositions that are sufficiently probable, a posteriori. p∈B EEU-maximization entails a normative Lockean thesis, with w threshold r+w . This explains (†), since w ≤ r would permit B(p) when b(p) ≤ 1/2, which is not in the Lockean spirit. Comparison Since our agents are (Bayesian) conditionalizers, this immediately suggests a natural diachronic requirement. The overall EEU of an agent’s belief set B, from the point of view of her credence function b(·) is defined as X EEU (B, b) Ö EEU (B(p), b) + AGM Revision We just explained how EUT implies a synchronic coherence constraint — the normative Lockean thesis. w∈W b(p) > EUT Revision 7 Shear & Fitelson Two Approaches to Belief Revision 8 Setup EUT Revision AGM Revision Comparison b(p | q) Future Work References Extras Setup EUT Revision Future Work References Extras AGM’s underlying principle is the principle of Conservativity (also sometimes called the principle of informational economy, or minimal mutilation). 0.8 0.6 Conservativity. When an agent learns E, she should revise to a posterior belief set B0 such that (a) B0 accommodates E, (b) B0 is deductively cogent, and (c) B0 constitutes the minimal change to B which satisfies (a) and (b). 0.4 0.2 0.6 0.7 0.8 0.9 1.0 Figure: The “degree” to which EUT belief updating can violate Leitgeb’s (P2), as a function of the EUT Lockean threshold t = Shear & Fitelson EUT Revision t= + Comparison The AGM revision operator can be characterized axiomatically. Next, we look at the AGM axioms. w r+w . 9 Two Approaches to Belief Revision AGM Revision AGM’s core synchronic requirement is Cogency (which needs to be assumed as a standing constraint) and its revisions are generally “logical” in nature. Future Work References Extras Shear & Fitelson Setup 10 Two Approaches to Belief Revision EUT Revision AGM Revision Comparison Future Work References Extras Given the other AGM axioms, Superexpansion and Subexpansion imply Inclusion and Vacuity, respectively, assuming only the following weak additional postulate. Basic AGM postulates: (*1) B ∗ E = Cn(B ∗ E) Closure (*2) E ∈ B ∗ E Idempotence. B ∗ > = B. Success (*3) B ∗ E ⊆ Cn(B ∪ {E}) Inclusion (*4) If E is consistent with B, then B ∗ E ⊇ Cn(B ∪ {E}) Vacuity (*5) If E is not a contradiction, then B ∗ E is consistent Consistency (*6) If X a Y , then B ∗ X = B ∗ Y Extensionality Supplementary AGM postulates: (*7) B ∗ (X ∧ Y ) ⊆ Cn((B ∗ X) ∪ {Y }) (*8) If Y is consistent with Cn(B ∗ X), then B ∗ (X ∧ Y ) ⊇ Cn((B ∗ X) ∪ {Y }) Shear & Fitelson Comparison The AGM theory of belief revision (Alchourron, Gärdenfors & Makinson [1]) is the most widely investigated and influential account of qualitative belief revision. b(p) = b(q) = t 1.0 Setup AGM Revision Two Approaches to Belief Revision Superexpansion Subexpansion 11 1. 2. 3. 4. 5. 6. Y ∈B∗X B ∗ X = B ∗ (> ∧ X) Y ∈ B ∗ (> ∧ X) Y ∈ Cn((B ∗ >) ∪ {X}) Cn((B ∗ >) ∪ {X}) = Cn(B ∪ {X}) Y ∈ Cn(B ∪ {X}) Assumption (1), Extensionality (1), (2), Logic (3), Logic, Superexpansion Idempotence, Logic (4), (5), Logic 1. 2. 3. 4. 5. X Y Y Y Y Assumption Assumption (2), Idempotence, Logic (3), Subexpansion (4), Extensionality Shear & Fitelson is consistent with B ∈ Cn(B ∪ {X}) ∈ Cn((B ∗ >) ∪ {X}) ∈ B ∗ (> ∧ X) ∈B∗X Two Approaches to Belief Revision 12 Setup EUT Revision AGM Revision Comparison Future Work References Extras Setup EUT Revision AGM Revision Comparison Future Work References (*4) If E is consistent with B, then B ∗ E ⊇ Cn(B ∪ {E}) Proof: We just need an example in which E is consistent with B, but B ÷ E É Cn(B ∪ {E}). To this end, consider the credence function generated by the following assignments: Closure (*2) E ∈ B ∗ E Success (*5) If E is not a contradiction, then B ∗ E is consistent Consistency (*6) If X a Y , then B ∗ X = B ∗ Y Extensionality (*7) B ∗ (X ∧ Y ) ⊆ Cn((B ∗ X) ∪ {Y }) Superexpansion (*8) If Y is consistent with Cn(B ∗ X), then B ∗ (X ∧ Y ) ⊇ Cn((B ∗ X) ∪ {Y }) (*9) B ∗ > = B Vacuity Proposition. ÷ does not satisfy Vacuity (or Subexpansion). Alternative Axiomatization of AGM, using Idempotence. (*1) B ∗ E = Cn(B ∗ E) Extras Subexpansion p b(p) E∧X E ∧ ¬X ¬E ∧ X ¬E ∧ ¬X .3 .05 .325 .325 Given a Lockean threshold of t = 0.9, the prior belief set is Idempotence B = {E ⊃ X}. And, the (EUT) posterior belief set is B0 = B ÷ E = {E, E ∨ X, E ∨ ¬X}. Shear & Fitelson Setup EUT Revision 13 Two Approaches to Belief Revision AGM Revision Comparison Future Work (*3) B ∗ E ⊆ Cn(B ∪ {E}) References Extras EUT Revision AGM Revision Comparison Future Work References Extras Because EUT satisfies Inclusion, it will never require the agent to gain more new beliefs than AGM. However, EUT may require the agent to lose more beliefs than AGM. Proposition. B ÷ E ⊆ Cn(B ∪ {E}) This feature (in conjunction with the fact that AGM may require the agent to gain more new beliefs than EUT) shows that EUT is less demanding of an agent (commitment-wise). Proof: Suppose X ∈ B ÷ E. Then, b(X | E) > t. And, it is a theorem of probability calculus that Pr(E ⊃ X) ≥ Pr(X | E). Therefore, b(E ⊃ X) > t. So, E ⊃ X ∈ B. Hence, by modus ponens (for material implication), X ∈ Cn(B ∪ {E}). In this sense, AGM is more risk-seeking in its diachronic requirements. Pettigrew [8] has independently argued (via thw use of an epistemic Hurwicz Criterion) that the insistance on Cogency is epistemically risk-seeking. A similar argument shows that EUT revision satisfies the more general principle Superexpansion. Suppose P ∈ B ÷ (X ∧ Y ). Then, b(P | X ∧ Y ) > t. It is a theorem of probability calculus that Pr((X ∧ Y ) ⊃ P ) ≥ Pr(P | X ∧ Y ). Therefore, b((X ∧ Y ) ⊃ P ) > t. So, (X ∧ Y ) ⊃ P ∈ B. So, by Success and modus ponens, P ∈ Cn((B ÷ X) ∪ {Y }). Two Approaches to Belief Revision Setup 14 Two Approaches to Belief Revision The crucial difference between AGM and EUT (aside, of course, from Cogency) is Vacuity/Subexpansion. Inclusion Intuitively, Inclusion requires that a revision does not include any more than the logical closure of the union of the original beliefs with the learned proposition. Shear & Fitelson Shear & Fitelson We close with a final theorem, which illuminates the tight connection between EUT’s violations of Vacuity (and Subexpansion) and its “risk aversion” (vs AGM revision). 15 Shear & Fitelson Two Approaches to Belief Revision 16 Setup EUT Revision AGM Revision Comparison Future Work References Extras Setup EUT Revision EUT violates Vacuity (wrt B, E) a E is consistent with B and Proof. Future Work Longer term, we also want to investigate iterated revision (EUT vs AGM), and also theories of belief update (in contrast to revision). We’re planning a series of four papers . . . 17 Two Approaches to Belief Revision Comparison Extras Jan van Eijck & Bryan Renne [5] recently provided a modal logic for belief given a Lockean threshold of 1/2. A near-term task is to investigate how their modal logic may be used to define a system of belief revision for a threshold of 1/2. (⇐) Suppose E is consistent with B and B ÷ E ⊂ B ∗ E. Then, there exists an X such that X ∈ B ∗ E but X ∉ B ÷ E. Because E is consistent with B, Vacuity and Inclusion imply that B ∗ E = Cn(B ∪ {E}). Therefore, X ∈ Cn(B ∪ {E}); but, X ∉ B ÷ E. AGM Revision References It would be nice to have a purely qualitative characterization/axiomatization of EUT Revision. Ideally, we’d like to have one for arbitrary Lockean thresholds. (⇒) Suppose EUT violates Vacuity (wrt B and E). Then, (a) E is consistent with B; and, (b) B ÷ E É Cn(B ∪ {E}). By (b), there exists an X such that X ∈ Cn(B ∪ {E}) but X ∉ B ÷ E. It follows from (a), Vacuity and Inclusion that Cn(B ∪ {E}) = B ∗ E. Therefore, X ∈ B ∗ E and X ∉ B ÷ E. And, by Inclusion, B ÷ E ⊆ Cn(B ∪ {E}) = B ∗ E. EUT Revision Future Work It would be useful to investigate EUT Revision from a “non-classical probability” perspective (e.g., two-place Popper functions, imprecise probability functions, etc.). B ÷ E ⊂ B ∗ E. Setup Comparison While we now have a clearer understanding of the (interesting) ways in which EUT and AGM revision diverge, there is much more work to be done in this project. Theorem Shear & Fitelson AGM Revision References Extras Shear & Fitelson Setup EUT Revision 18 Two Approaches to Belief Revision AGM Revision Comparison Future Work References Extras [1] Carlos Alchourron, Peter Gärdenfors, and David Makinson. On the logic of theory change: Partial meet contraction and revision functions. The Journal of Symbolic Logic, 50(2): 510–530, June 1985. [2] D. Christensen, Putting Logic in its Place, OUP, 2007. 1. 2. 3. 4. 5. 6. 7. 8. 9. [3] Kevin Dorst. An Epistemic Utility Argument for the Threshold View of Outright Belief Manuscript, 2014. [4] Kenny Easwaran. Dr. Truthlove, or how I learned to stop worrying and love Bayesian probabilities. Manuscript, 2014. [5] Jan van Eijck and Bryan Renne. Belief as Willingness to Bet. Manuscript, 2014. [6] Hannes Leitgeb. The review paradox: On the diachronic costs of not closing rational belief under conjunction. Noûs, 78(4): 781–793, 2013. [7] David Makinson and James Hawthorne. Lossy Inference Rules and their Bounds: A Brief Review. The Road to Universal Logic: Festschrift for 50th Birthday of Jean-Yves Béziau, Vol. 1, pages 385–407, Springer, 2015. [8] Richard Pettigrew. Accuracy and Risk: a Jamesian investigation in formal epistemology. Manuscript, 2015. B is consistent. B is closed, i.e., B = Cn(B). X ∈ B. X is consistent with B. B ∗ X = Cn(B ∪ {X}). B ∗ X = Cn(B). B ∗ X = Cn(B ∗ X) Cn(B ∗ X) = Cn(B) B∗X =B Assumption Assumption Assumption (1), (3), Logic (4), Vacuity, Inclusion (5), (3), Logic Closure (6), (7), Logic (7), (8), (2), Logic Figure: Derivation of (P2) from Closure, Inclusion, and Vacuity 1 [9] Alexander Pruss. The badness of being certain of a falsehood is at least log(4)−1 times greater than the value of being certain of a truth. Logos & Episteme, 3(2): 229–238, 2012. [10] Florian Steinberger. Explosion and the normativity of logic. Mind, to appear. Shear & Fitelson Two Approaches to Belief Revision 19 Shear & Fitelson Two Approaches to Belief Revision 20 Setup EUT Revision 1. 2. 3. 4. 5. 6. AGM Revision Comparison Y ∈B∗X B ∗ X = B ∗ (> ∧ X) Y ∈ B ∗ (> ∧ X) Y ∈ Cn((B ∗ >) ∪ {X}) Cn((B ∗ >) ∪ {X}) = Cn(B ∪ {X}) Y ∈ Cn(B ∪ {X}) Future Work References Extras Assumption (1), Extensionality (1), (2), Logic (3), Logic, Superexpansion Idempotence, Logic (4), (5), Logic Figure: Derivation of Inclusion from Superexpansion, Extensionality, and Idempotence Setup EUT Revision AGM Revision p b(p) E∧X E ∧ ¬X ¬E ∧ X ¬E ∧ ¬X E X E≡X E ≡ ¬X ¬E ¬X E∨X E ∨ ¬X ¬E ∨ X ¬E ∨ ¬X .3 .05 .325 .325 .35 .625 .625 .375 .65 .375 .675 .675 .95 .7 Comparison b(p | E) 6/7 1/7 0 0 1 6/7 6/7 1/7 0 1/7 1 1 6/7 1/7 Future Work p ∈ Cn(B ∪ {E})? Yes No No No Yes Yes Yes No No No Yes Yes Yes No References Extras p ∈ B ÷ E? No No No No Yes No No No No No Yes Yes No No Table: Full counterexample to Vacuity for EUT Revision Shear & Fitelson Setup 21 Two Approaches to Belief Revision EUT Revision AGM Revision p b(p) E∧X E ∧ ¬X ¬E ∧ X ¬E ∧ ¬X E X E≡X E ≡ ¬X ¬E ¬X E∨X E ∨ ¬X ¬E ∨ X ¬E ∨ ¬X .3 .05 .325 .325 .35 .625 .625 .375 .65 .375 .675 .675 .95 .7 Comparison b(p | E) 6/7 1/7 0 0 1 6/7 6/7 1/7 0 1/7 1 1 6/7 1/7 p ∈ B? No No No No No No No No No No No No Yes No Future Work p ∈ B ÷ E \ B? No No No No Yes No No No No No Yes Yes No No References p ∈ B ∗ E \ B? Table: Demonstration that B ÷ E ⊂ B ∗ EB in our example Shear & Fitelson Two Approaches to Belief Revision Extras Yes No No No Yes Yes Yes No No No Yes Yes No No 23 Shear & Fitelson Two Approaches to Belief Revision 22