Jeffrey - Probability Kinematics

Premisse: any adequate account of an agent's attitudes (degrees of beliefs) must include how they can change over time.

Jeffrey discusses attitudes in terms of the "probability" and "desirability" of a certain proposition.

I do not know how Jeffrey understands "desirability", because it refers to preceding chapters that I did not read.

He writes that des AB is prob AB / prob B which is indeed the definition of conditional probability prob (A|B).

This chapter is not necessarily about the desirability part so I'll just ignore that.

Limits of conditionalization

Conditionalization occurs on observation of evidence E. The new subjective probability is thus warranted if evidence-proposition E is indeed applicable.

Problem: sometimes a belief update is warranted whereas it cannot be clearly stated that some evidence-proposition E is true.

Observation by candlelight

What if there is uncertainty about the precise content of E?

Imagine observing some cloth by candlelight and conceding the cloth might be green, blue or (improbably) violet. What proposition E (that we consider to be true) captures this experience so that we may condition upon it? Your beliefs would certainly be updated w.r.t. your prior beliefs. A solution could be to have E describe the exact experience, but it is not given that such a proposition would be possible, and in any case our language would fall short and be too vague.

Nor is it a solution to try and be exact, in describing the experience in physicalist terms (e.g. stimulations, rods, cones etc.). Not even the best physicist could describe his own experience in this manner.

Is this a problem?

Well, when people such as C.I. Lewis take strict conditionalization as the only way of assimilating evidence and updating one's beliefs, then you need to resort to weird constructions stating that there must be some "expressive" language expressing the exact contents of the evidence E.

Jeffry makes two points, I think:

  1. Bayesianism is normative and not descriptive, and thus does not have to commit to being 100% descriptively accurate.

the theory of preferences is ours, but not necessarily the agent's. The theory can be true of an agent even if he is quite ignorant of the theory, and in particular, the language in which we express the propositions in his preference ranking need not be a language that the agent understands.

In other words, Bayesianism is not descriptive (human beings are no de facto Bayesians) but normative for human decision making.

  1. The philosophical conundrums about some expressive language in which the evidence-proposition E is true, so that we may appropriately apply conditionalization, may be avoided simply by having an update rule for uncertain evidence.

Reasoning behind Jeffrey's Rule

The question that is to be answered is how exactly subjective belief in proposition A has to be updated after a "passage of experience" with regard to B (that does not directly already influence A in and of itself), where there is uncertainty in the evidence B.

Start with binary events.

The first important insight (11.1 in the paper) is that Pr(A|B) does not change when B changes.

Jeffrey shows this with the mudrunner example: You know a racehorse that runs extremely well on a muddy track. You now hear a weather forecast that predicts that it will likely rain, i.e. the track is likely to be muddy. A describes the probability of winning, B describes the probability of a muddy track.

This warrants two conclusions:

  1. Now that Pr(B) is higher, you should be more certain in your belief that the horse will win Pr(A).
  2. But this change in Pr(B) does not warrant a change in Pr(A|B)!

In the somewhat confusing notation of the paper, point 2 is expressed as PROB(A|B)=prob(A|B) (and idem for the negation of B, in case of a binary event), where the capitalized version indicates the updated belief.

My attempt at deriving Jeffrey's rule from this insight:

PROB(A) = PROB(A,B)+PROB(A,~B).

That is, we use the law of total probability.

Now, we know that we can rewrite that using Bayes rule to:

PROB(A) = PROB(A|B) * PROB(B) + PROB(A|~B) * PROB(~B).

Now, based on the reasoning above, we know that the new updated conditional probability PROB(A|B) is in fact Prob(A|B), i.e. the same conditional probability. This means that we can express the new subjective probability of A, solely in terms of the new distribution of belief (in this case still binary) over B!

AMAZEBALLS.

So:

PROB(A) = prob(A|B) * PROB(B) + prob(A|~B) * PROB(~B).

So in this way, we can express uncertainty in the evidence used in the Bayesian update rule.

I've used this rule many times, but never quite understood this important difference between which value is "old" and which one "new".

Relevance measure

TODO.