Bayesianism I: Introduction and Arguments in Favor | by Easwaran, Philosophy Compass 6/5 (2011), 312-320
@Bayes
What is common to everything known under the label "Bayesianism"?
Mainly, representation of belief as coming in degrees. There are much varieties in understanding what Bayesianism is.
N.B. remarks and study questions from the syllabus are labeled with Q.
Core ideas (mostly agreed on, p. 312)
- There is an mental attitude of degree of belief or credence and we can often assign this credence a numerical value
- Central claim, but is degree of belief the only important "doxastic and epistemic attitude"?
- For rationality an agent's degree of beliefs must follow the axioms of probability theory
- Synchronically: Has to hold at any single time
- Conditionalization (or something related) is the default way beliefs change over time.
- Diachronically: describes how rationality is preserved over time
- I.e. you come to know new information and this changes your beliefs
Bayesianism and probabilism
- Full Bayesianism: at least 1-3
- Probabilism: excluding 3, so 1-2.
Both agree on the usefulness of mathematically representing degrees of belief (core idea 1 and 2). How can we express we are more sure of our belief about the sun rising that it raining tomorrow?
Philosophers debate the adequacy of this representation itself (so even may question idea 1).
Q: For a discussion of what numerical values should be ("where do the probabilities come from"), see the Fiorillo paper on Jaynes.
Kolmogorov's axioms of rationality (w.r.t. subjective probabilities)
Cf. axioms of probability theory.
Have a function P defining assigning real numbers to all members of F, where F is a collection of subsets of the "field of probabilities" E.
Axioms on P:
- values of P are non-negative
- P(E)=1: field of all probabilities describes all possible outcomes, and thus at least one of those outcomes must occur (my phrasing)
- If A and B have nothing in common, then $P(A \cup B) = P(A) + P(B)$.
Q: Kolmogorov axioms are defined on sets of events. So probability theory satisfies the "normal" axioms of set theory. The three axioms of Kolmogorov together with axioms of set theory fully define probability calculus.
$\Omega = {1,2,3,4,5,6}$
$A \subseteq \Omega = {1,2,3}$
Then set theory states $B \ A = {4,5,6}$
Together with Kolmogorov's 2nd and 3rd it follows Pr(B)=1-Pr(A).
Conditional possibility: mathematical definition or pre-theoretic notion?
Hàjek (2003) argues that conditional degree of belief is a pre-theoretic notion as much as a degree of belief, and therefore we must analyze it, rather than just defining it mathematically. (313)
Also note that unlike the other notions, this one does not follow from the Kolmogorov set definitions!
A strictly mathematical definition has some issues with zero-probabilities.
- B&A entails A.
- So P(A) must at least be as big as P(B&A), but can be bigger (since A does not entail B&A)
- But if P(A)=0, 2 implies that P(B&A)=0, and thus we also know that P(B|A)=0/0
Interesting remarks:
- Nothing is specifically Bayesian about Bayes' rule, as it does not necessarily imply the notion of degrees of belief
- Neither does it describe updating beliefs (see idea 3), but it is strictly synchronic.
- So another argument is needed to show the relevance of Bayes' rule for Bayesianism (Easwaran will do so in relation to "Bayesian confirmation theory")
How to relate these axioms to the epistemic interpretation of belief?
Interpretation:
- E is the set of epistemic possibilities, i.e. everything we may know
- Can we indeed express propositions (about which we may have a belief) in terms of sets?
- What are epistemic possibilities and do they exist (as a set)?
Q: What is the difference between epistemic and alethic possibilities in modal logic?
Alethic possibility: deals with how the world will be or might have been.
Epistemic possibility: deals with subjective knowledge about the world.
"Bernie Sanders may have won the 2016 US elections": false, but if one does not know about the outcome of the 2016 elections this would be an epistemic possibility. But it is logically false and thus not an alethic possibility.
Subtle difference with "Bernie Sanders might have won the 2016 US elections": this is an alethic possibility; there is indeed a possible counter-factual world where this hold. But we know this is not what happened, so this is not an epistemic possibility.
Conditionalization: diachronic rules for rationality
Most basic rule for updating: conditionalization.
- Triggered by learning event, e.g. observation
- Conditioned upon new strongest learned proposition, update P_new(B)=P_old(B|A)
- If A is non-zero in P_old, it will also be so in P_new and satisfy the Kolmogorov axioms
- If P_new(A)=1, no more updating can occur.
Note that this diachronic conditionaliziation is not simply Bayes' rule.
To solve issue 4 (see problem with old evidence in the follow up paper), often Jeffrey updating is used because the evidence will never be 0 or 1.
P_new(B)=P_old(B|A)P_new(A) + P_old(B|not A)P_new(not A)
This is what we always used in previous classes as well, right? Why did no one explain this to me?
Issues with Jeffrey updating:
- Diaconis and Zabell (1982): this can be seen as a version of traditional conditionalization when adding particular new propositions to the language (?)
- Field (1978) and Wagner (2002): what does it mean to receive "the same evidence" in two different Jeffrey updates?
- receiving the same evidence in two different orders should result in the same final credences. Does this hold?
Other ways of belief updating
- Forgetting, loss of track of time, Arntzenius (2003).
- Sleeping Beauty Problem (Elga 2000)
- distinction possible worlds and epistemic possibilities?
- is self-locating information revelant to non-self-locating beliefs?
- Responses:
- Lewis (2001), Hitchcock (2004), Bradley and Leitgeb (2006), Meacham (2008), Titelbaum (2008).
- Dr. Evil Puzzle Elga (2004)
- Response Weatherson (2005)
It looks like you need a enigmatic or Jewish name in this field of science.
Arguments for Bayesianism
Dutch book argument
Most prominent pro-argument, attributed to Ramsey (1926) and de Finetti (1937).
Relies on assumptions about mental states of agents for showing that
- rational agents must conform to axioms of probability theory (idea 2)
-
must update by conditionalization (idea 3)
-
Let's say that for any proposition there is a price that an agent considers fair to a bet rewarding 1 dollar if true and nothing if false.
- Fair means that an agent is willing to sell the bet for more than the fair price, and buy another bet for less than the fair price
- Big assumption: willingness to buy or sell completely independent from other ongoing bets (obviously does not hold)
The Dutch book theorem: if Kolmogorov axioms do not hold, the agent is willing to accept a finite set of bets to which the agent is "guaranteed" to lose money. If they do, not (but still susceptible to a generalized Dutch book, whatever that means, cf. 315).
The notion of "guarantee" must be subjective. It's not the case that we know A is false, in face of which indeed betting on A is guaranteed to lose money.
The matter at hand is: if the agent has a relies on a set of epistemic possibilities where A is likely true, then this agent cannot guarantee she will lose money. There is a chance, from her perspective, she might win; this makes this bet rational from a subjective point of view.
This argument is supposed to show that a rational agent must satisfy Kolmogorov's axioms. (I do not get the details yet).
The main point is as well: this "fair price" is a way of assigning a numerical value to a degree of belief and of connecting Bayesian math with the epistemic Bayesianism.
But if an agent had a uniform fair price for buying and selling, and if this price were determined solely by her degree of belief in the proposition, then these assumptions would be true, and furthermore the fair prices would provide a way to assign numerical values to degrees of belief.
So the Dutch book argument can be seen as "the first claim of Bayesianism together with a sort of 'package principle' requiring that an agent be willing to buy a set of bets if she is willing to buy every bet in the set." (315)
Modified version by Christensen
Without the full unrealistic assumptions of the argument.
1) "we can just assume that there is a Bayesian notion of degree of belief, and that this notion justifies the agent in deeming bets favorable (regardless of whether she would be willing to accept them or not)" (315) 2) "Then we assume that an agent is justified in deeming a combination of bets favorable if she is justified in deeming each one individually favorable" (315) 3) and "that a rational agent is never justified in deeming favorable a combination of bets that guarantees her loss."
Conclusion:
The Dutch book theorem shows that an agent will violate Kolmogorov's axioms iff there is a Dutch book, each of whose bets she is individually justified in deeming fair, and thus a rational agent will never violate Kolmogorov's axioms.
Waarom? Ik volg dit nog steeds niet helemaal? Ik snap het tot de allerlaatste zin.
Dutch book variety for conditional probability
Bet A|B pays out if A & B are true; pays nothing if A is false but B is true; refunds if B is false.
If then we assume that the agent's conditional degree of belief P(A|B) specifies the fair price for this bet, then she is vulnerable to a conditional Dutch book iff she violates P(A|B)P(B) = P(A&B). (316)
Diachronic Dutch books
Highly problematic and complicated. See 316.
Main figure is Teller (1973)
Q: How realistic do you think the prerequisites of Teller are in reality (the agent known in advance that she will learn which (of exactly one) piece of evidence is true)?
I did not really get Tellers reasoning. Review p. 316.
Reflection: One interesting issue on diachronic updating is "reflection", introduced by van Fraassen (1984).
an important constraint on updating known as 'reflection', one version of which states that if an agent is certain about what her future degrees of belief will be, then she should already have those degrees of belief. (316)
TODO: Review. I think this is very complicated, and it's explained on a high-level.
Summary
Dutch book arguments have been used to support the synchronic and diachronic claims of Bayesianism, in addition to helping give meaning to the primary claim that belief can come in degrees. (316)
Main problems:
However, the assumptions about fair betting prices that are needed seem unduly strong to many people. Additionally, the inconsistency they point out is an inconsistency of preferences, and not of beliefs, so many are worried that this can't be an argument for an epistemic conclusion. (316-317)
Representation theorems
No longer assumption that agents have fair betting prices. Instead, make assumptions that an agent has a preference ranking over all actions. (All including highly unrealistic ones).
- Structural assumption: agent has preferences among all "actions" that can be defined by a function from epistemic possibilities to "outcomes"
- Does not matter if those actions are really available, you just assume a ranking over all actions anyways.
- Transivity
- Dominance
Given these assumptions, there is a unique function P that indicates for two propositions A and B which one is preferred. See 317.
Then the theorist makes one final leap and says that U represents the agent's 'utilities' (how desirable she finds various outcomes) and P represents her probabilities (how believable she finds various propositions). This identification is supported by the intuitive fact that preferences govern actions just as desires and beliefs do, and that the calculation [formule expected utility here] is the standard 'expected utility' used in decision theory to represent the notion of rational preference between actions.
In the summary from the syllabus (Kwisthout):
The representation argument, in a nutshell, argues that for a particular set of assumptions and a (complete) set of preferences over outcomes O of a particular agent, tehre is a function U from the outcomes to R and a function P from propositions to [0,1] that links the agent's preferences to a utility function U and probability measure P. (24)
Counter arguments focus mainly on limited expressiveness of utility theory. See 317-318.
Interesting:
Although the theorem doesn't directly assume the existence of fair prices for bets, standard decision theory guarantees that these fair prices exist, so worries about Dutch book arguments can often be translated into worries about decision theory. (318)
And!
representation theorems don't directly put any constraints on the diachronic change of degrees of belief, so they are at best a partial vindication of Bayesianism.
Other arguments and closing remark
one particular feature that all these arguments seem to share is the fact that they derive rational constraints on degree of belief from apparently pragmatic constraints on preference. This is seen by many to be a failing, since epistemic norms and practical norms seem to be two different sorts of things. (318)
There is thus a need for non-pragmatic approaches, e.g. Jaynes 2003.