Bayesianism 2: Applications and Criticisms | Kenny Easwaran
Great text. Continuation of Easwaran - Bayesianism I .
Carnap argued that probability gives the best notion of confirmation.
Carnap sought not just a qualitative theory of whether a given piece E of evidence confirms, disconfirms, or is neutral with respect to a hypothesis H, but also a theory of when E1 confirms H1 more strongly than E2 confirms H2 (321)
But Carnap failed to distinguish between
- an absolute notion of confirmation, a.k.a. "firmness": P(H|E)
- an incremental notion for 'increase in firmness': P(H|E)-P(H)
In general, the Bayesian approach favors the latter. (I would add that Bayesian theory often discusses hypotheses in a contrastive setting).
Standard Bayesian confirmation theory with a qualitative notion of confirmation:
E confirms H iff P(H|E) > P(H), i.e. a scientist has some prior degree of belief in a hypothesis. She does an experiment where E is learned. Did the confidence in the hypotheses increase after learning E P(H|E)? If yes, then an experiment confirms the expected hypothesis (it confirms the expectation of the experiment).
Note that Bayes' rule implies P(H|E) > P(H) <-> P(E|H) > P(E).
There's plenty of disagreement about the interpretation of probability though.
Terminology
All the usual, but Easwaran makes a remark about something I only figured out myself recently after being very confused about it.
In Bayesian confirmation theory, we speak of the "likelihood" of an hypothesis. But whereas the likelihood of the data in Bayes rule is P(E|H), the likelihood of not P(H|E) but also P(E|H). (Cf. 322) That is: a good hypothesis is one that makes the data more likely.
Confirmation measures
Above we expressed the increase in firmness as P(H|E)-P(H). An obvious limitation of this measure is that you can't actually gain any confirmation of an hypothesis when the prior is already high.
Many confirmation measures disagree about the strenght of confirmation but in general agree on whether something is confirmation, disconfirmation or irrelevant.
- P(H|E)-P(H|~E)
- P(E|H)/P(E|~H)
- P(H|E)/P(H)
Is this pluralism of confirmation measures problematic?
On the one hand, yes, because they can differ a lot in the degree of confirmation. On the other hand, no, because all these measures have their respective strengths for capturing particular intuitions.
Popper-Miller argument
The main objection is that while positive deductive support is possible, there can be no positive inductive support.
It's a bad argument, basically, see p.323.
The takeaway is that deductive support can not always easily be told apart from inductive support. But the next quote illustrates the difference lucidly:
The observation of a given raven as being black gives some minor deductive support to the hypothesis that all ravens are black (by showing that one potential deductive falsifier doesn't falsify it), but the more interesting question is what inductive support it gives to the claim that other ravens are black. (323)
Logical Omniscience
This is a bigger issue.
Any logical statement is by definition tautological and thus necessarily true. Now, standard interpretations of probability axioms usually imply that logica truths should be assigned a probability of 1. However, in practice people are in fact often uncertain about logical truths! "consider logic students who don't recognize a claim as a tautology, or mathematicians wondering about whether a given conjecture follows from the axioms." (232)
You could change the interpretation of Kolmogorov sets, but then the relation with syntactic axiomatizations such as our usual probability theory becomes problematic.
Some ideas:
Garber 1983:
- An agent uses some language
- But the agent's credences do not have to be expressed in the same language
- If we treat each sentences in the agent's language as an atom in the theorists language, then we can define the probability axioms for that language.
- All agent sentences, even tautologies, are then contingent sentences in the theorists language
Downsides:
- You still have logical omniscience in the theorists language
- Point I do not get: "why the theorist's language omits the logical structure of the agent's language, even where the agent recognizes it."
Gaifman 2004: This answers is too abstract for me to make much sense.
- Preserve agent language, but "gives a different syntactic axiomatization of probability theory" (324)
- under these axioms "the agent is only required to be logically omniscient about statements that have a deductive proof, every one of whose steps is in the designated fragment." (324)
- logical omniscience is avoided because even though one step may be graspable, all the necessary proofs to reach it may not be immediately graspable.
Another solution: saying that Bayesian confirmation theory concerns ideal agents, and that it's fine and to be expected that actual scientists are non-ideal. This answer could be problematic for some arguments that rely on Bayesian theory not implying logical omniscience.
Old evidence
If we have existing evidence with P(E)=1, then it cannot be used to confirm any (new) hypothesis if we measure confirmation as P(H|E)-P(H) because then P(E|H)=P(H).
Standard example: the perihelion shift of Mercury was already known to Einstein when he realized his own theory explained it better than the Newtonian one. So this was further support for Einsteins theory. However, since he already knew (and did not doubt) the value of the perihelion shift, the standard Bayesian account instead says that this old evidence could not have confirmed his hypothesis. This is obviously undesirable.
Garber 1983: "concede that the perihelion shift itself gave no confirmation for Einstein's theory, but to claim that the fact that Einstein's theory entailed the perihelion shift gave the confirmation" (324).
So then the evidence is not the old evidence, but the evidence is the finding of new logical information, namely learning that Einstein's theory entails the perihelion shift.
This works, but because this argument uses logical information, it needs to solve the problem of logical omniscience (but it needed to do so anyways). If you don't solve logical omniscience, then the "new" logical information also cannot be new.
My question here: the lines between induction and deduction really blur here.
Old new evidence
For people after Einstein, this logical information that was new to Einstein is again old for them and also has P(E)=1.
This can be called the problem of old new evidence: that of evidence that is old but was relevant when it was new.
Einstein's problem, phrased in these terms, was that of new old evidence: old evidence that is newly recognized as relevant to a theory.
Q: Can you explain the difference between old new evidence and new old evidence?
How can Bayesian theory deal with these nuances?
Proposed solutions
It's not useful to try to go back to the original moment where the evidence was new to get a sensen of its confirmation, because it's hardly useful to do this when later developments made the evidence irrelevant, for example. Instead, we need to express confirmation w.r.t. the present. (325)
-
Do not use agent's actual belief where P(E)=1, but use the degree of belief that "the agent would have had, if she hadn't already known E". It is however unclear if such a function would be unique and how we would be able to find it. Also, it is unclear what we do if such a function would differ drastically from the actual belief function.
-
Williamson (2002) and Carnap (1960) have an extreme version of this where they suggest that P is some sort of hypothetical prior that an agent would have if it would have no empirical knowledge at all. But this is even more problematic.
Main objection: "this suggestion seems to ignore the fact that much evidence is only evidence in light of our other background knowledge" (325).
- Endorse Jeffrey conditionalization instead of strict conditionalization. In this case, P(E) will not be 1. Then you deal with the previous issue that the degree of confirmation will be very little. But in the pluralistic account, there are at least some measures who can still indicate some more confirmation when the prior on the evidence is very high.
A more fundamental problem here: "this still seems to make the evidential relations already holding between an agent's beliefs dependent on how future increases in P(E) would affect P(H), which seems to get something conceptually wrong." (325)
Q: I do not follow this. I should review Jeffrey conditionalization first.
New theories
Bayesian theory also does not really account for new theories.
In a sense, the logical omniscience assumed in Bayesian theory goes beyond assigning probability 1 to tautologies, and also requires assigning probabilities to every sentence whatsoever, regardless of which ones the agent may have actually ever considered. (325)
Old evidence & new theories. Intuitively, there should be a difference in:
a. "an evidential relation between an old piece of evidence that a new hypothesis happens to explain" (326) b. and "an old piece of evidence that a new hypothesis was specifically designed to accommodate" (326)
Then the contention is that "somehow the logical entailment relation between the hypothesis and evidence seems like it should be much more relevant in one case than the other, even though the theory had no prior before it was thought up. (326)"
Q: Actually it's not that clear to me which one should have a stronger confirmatory relation.
In any case, these nuances are hardly explained by conditionalization, especially when priors are missing.
Problem of the priors
The degree of confirmation is highly dependent on the choice of priors.
Why?
Consider Bayes rule: $P(H|E) = \frac{ P(E|H) }{ P(E) } x P(H)$.
We already said that the degree of confirmation depends on $P(H|E)-P(H)$. The "amplification" of P(H) only depends on P(E|H)/P(E).
The problem is that in some optimal cases we can relatively objectively specify P(E|H), but the choice of the prior P(E) is more problematic. As you can see, the amount of confirmation does heavily depend on it. The more certain P(E), the lower the amount of confirmation.
Are multiple priors permissible?
If yes, then two scientists can have two different priors and can reach different degrees of confirmation despite seeing the same evidence.
To avoid this, objective Bayesians prescribe that there should be (in the strongest version of this objective strand) one rational set of prior probabilities.
They apparently prefer the Cox argument as a justification of Bayesianism.
Q: What is the Cox argument?
Objective Bayesians often subscribe to the Principle of Indifference: "rational agents should divide their credences equally among possibilities if they have no evidence favoring one over another" (326)
The Bertrand Paradox for the principle of indifference
Subjective Bayesians would point out that in fact the Principle of Indifference does not straightforwardly propose a unique prior distribution.
Version from Van Fraassen:
- we have a cubical box width a side length between 1 and 3 meters
- so we assign 1/2 credence to it being 1-2 meters, and 1/2 credence to it being 2-3 meters (that is a uniform distribution)
- we can translate this problem into an equivalent estimate, namelijk that the side surface will be between 1 and 9 square meters.
- then if we maintain a uniform distribution, we end up with 3/8 credence for 1-4 and 5/8 credence for 4-9.
So we have in fact assigned two different credences to equivalent statements.
TODO: review Jaynes' answer. See slides.
Subjective Bayesians's alternative
Subjective Bayesians must instead show that, when they allow priors to differ, that in general scientist that are repeatedly confronted with the same evidence will have converging credences.
In other words, "that scientists generally tend to agree as to which pieces of evidence confirm which hypotheses, since different priors will often give different confirmation relations." (327)
Bayesian Epistemology
The radically new element is that belief is no longer all of nothing, but comes in degrees.
The question is: how do degree of beliefs map onto existing notions of belief, justification, knowledge.o
Do we have P(X)=1 indicate a full belief that is unrevisable? Or do we have "non-extreme degrees of belief for everything" (327) This is Jeffrey's position of radical probabilism i.c.m. with Jeffrey's updating.
Goodman's riddle
The requirement that H have a non-zero prior also turns out not to be innocuous, because it is just this assumption that prevents a problem with Goodman's 'new riddle of induction'. It turns out that if n observations of green emeralds are sufficient to give the agent credence 0.99 that the next emerald will be green, then she must have initially assigned credence at most 0.01 to any hypothesis of the form 'all emeralds are grue_m' for m > n, where 'grue_m' means 'green and one of t he first m, or blue one of the later emeralds'. Induction is saved without getting an outright contradiction from the new riddle, but the question of how to actually deal with grue is pushed into the priors. (328)
So if you have a non-zero prior, instances of a universal prior (all X are Y etc.) will give "each other arbitrarily strong support" (328)
Why?
I do not get Easwaran's proof on page 328. He says that when you have some universal prior H and a sequence of observations that are logically entailed that the posterior credence in P(H|E1,...En) = P(H)/P(E1, ..., En). But I do not see how this is the case.
Got it! A very important detail here is that all observations are logically entailed by the universal prior H. This means that P(E1,...,En|H)=1, so it drops out of Bayes' formula.
His main conclusion is that the sequence of evidences should converge to 1. But we now can conclude that this has to be the case, otherwise the posterior would end up being higher than 1!
This would be a solution to Goodman's riddle because this basically is an account of positive inductive support that shows that the sensible hypothesis "the next emerald is green" will have tremendously high support whereas the awkard hypothesis doesn't.
TODO: I'm still not entirely sure how this solves Goodman's riddle though.
The Lockean thesis
Lockean thesis: about some threshold t, we can speak of a full belief.
Lottery paradox of Henry Kyburg:
If there were some threshold t < 1 such that having a degree of belief at least t suffices for full belief in a proposition, then we can consider a fair lottery with more than 1/(1-t) tickets, of which the agent has exactly one. She ougth to believe to degree greater than t that this ticket is not a winner, and thus the Lockean thesis says that she ought to believe her ticket will not win. However, many people share the intuition that in these sorts of cases, she doesn't know that the ticket won't win, and therefore she ought not to believe that. (328)
A solution of setting t to 1 is undesirable, because then it seems we never have full beliefs.
Vs. frequentist statistiscs
Section on the differences between Bayesian and frequentists statistics. The explanation is very nice but there's no need to record it here.
Interesting is the observation that these two approaches are, in practice, not as clearly separated as they are theoretically.
In practice however, the different methodologies aren't so cleanly separated. Frequentists often do gain confidence from small p-values, even though their interpretation of their procedure gives them no justification for this. Bayesians often do accept or reject hypotheses, even though their method gives no bound on the chance of making mistakes. It seems that resolving the debates will require attention to the relationship between the attitudes of degree of belief and acceptance or rejection of hypotheses, as well as sensitivity to the actual goals that scientists want to achieve by means of statistical testing. (330)