Wednesday, May 27, 2015

Reflection and Sleeping Beauty


I’ve been reading through a lot of early papers on the Sleeping Beauty problem recently. (Check out SleepingBeautyProblem.org!)One of the main issues is how the problem relates to van Fraassen’s Reflection Principle. The principle states that your present probability assignments should be the same as your known future probability assignments (or at least lie within the range that spans what you take your future assignments to possibly be).  So if you know that you’re going to watch a debate tomorrow on whether P, and you’re going to come out thinking that probably P, you should already think that probably P.

Moreover, your present probabilities conditional on having a given future probability should be equal to those future probabilities:

Reflection: Pr(A|pr+) = Pr+(A)                      where ‘pr+’ = Pr+() is your future probability assignment

Reflection has its faults, but there is something intuitively right about it.  It is odd for there to be foreseeable rational changes in assignments – that is, it is odd if there is some kind of evidence that I could get, such that I know that I am going to get it and that it’s going to convince me of something, and not be convinced of it already.

On the most popular answer to the Sleeping Beauty problem, Thirdism, the Reflection Principle is violated. (It also violated by Lewis’s version of Halfism). At some point Beauty should have some foreseen change of probability – so Beauty knows that while she assigns ½ to HEADS on Sunday, she knows she will later assign ½ (or ¾, for Halfers) to HEADS on Monday after awakening  (or after being told that it is Monday). 

Brad Monton suggested failure results from Beauty forgetting her self-locating information. Van Fraassen’s statement of the Principle seems to run into problems when we might forget things we had known previously. If you know that you’re not going to remember what you ate for breakfast in a year, that doesn’t mean you anything about what you should believe now. Beauty does lose some kind of information. So Monton’s suggestion makes some sense.

However, I think that Monton missed an important part of the story. There is something more substantive that can be said about why Reflection fails in this case, and the failure highlights an important lesson about deferential norms of probability.

A deferential norm of probability tells you what probability to assign to a given proposition in light of what you know about certain other special probability assignments. There are three main deferential norms of probability, though the equal weight view in the literature on disagreement might be seen as a fourth. The first is the Principal Principle. It tells you to assign probabilities that align with known chances. If you know a coin toss will land heads with a 50/50 chance, you should assign a 50% probability of it doing so. The second is the Principle of Reflection, which we’ve already seen. The third is the Rational Reflection principle. It tells you to assign probabilities in line with known rational probabilities. If you believe that the rational probability to assign to reality of global warming is .8, you should assign a .8 probability.

               Principal Principle: Pr(A|ch) = Ch(A)                                          where ‘ch’ = Ch() is the chance function

               Reflection: Pr(A|pr+) = Pr+(A)                      where ‘pr+’ = Pr+() is your future probability assignment

               Rational Reflection: Pr(A|rp)  = RP(A)        where ‘rp’ = RP() is the rational probability assignment

The Principal Principle is known to face problems when paired with certain reductive of objective chance. Suppose (for an over-simplified example) that you reduce facts about chances to frequencies: a coin has a ½ chance of landing heads if it lands heads half the time. Then given that the coin has a 50/50 chance of landing heads and will only be tossed twice, we can conclude that it will land heads once and tails once. If it has already landed heads, we know it will land tails. The Rational Reflection principle leads to interesting difficulties in the clock puzzle.

There is a similar fix for both problems (see Elga, “The Puzzle of the Unmarked Clock and the New Rational Reflection Principle” and Hall, “Correcting the Guide to Objective Chance”), that properly understood will help us understand why the Principle of Reflection fails in the Sleeping Beauty case.  

This fix is to make the probability assignments conditional:

Prinicipal Principle: Pr(A|ch) = Ch(A|ch)                    where ch is the statement that Ch( ) is the chance function.

               Rational Reflection : Pr(A|rp) = RP(A|rp)                  where rp is the statement that RP( ) is the rational probability assignment.

The general reason to be deferential to these probability assignments is that they are more likely to be accurate. There is no reason to withhold what we know from those whose judgment we should defer to. If we know something about the chance function, we shouldn’t seclude it from the chance function. If we know something about the rational probability function, we shouldn’t keep it secret from the rational probability assumption.

This carries over to the Reflection Principle. It doesn’t make sense to align one’s probabilities with ones known future probabilities if one presently knows more than one will in the future. Instead, we should align our present probabilities with known future probabilities conditional on what we now know. And moreover, what we assign in A conditional on some probability assignment T being worthy of deference should equal T given that T is worthy of deference.

This may not be thought to help explain why Reflection fails in Sleeping Beauty. After all, doesn’t Beauty know on Monday everything she knew on Sunday? In one sense, this is correct: she doesn’t actually forget anything between Sunday and Monday.

However, on Sunday Beauty knows more about her foreseen situation-on-Monday than she does on Monday. On Monday, she doesn’t know that it is Monday. On Sunday, she knows that the situation that she foresees that will assign different probabilities than she presently assigns will be on Monday. 

 This isn’t something that she forgets. But she later is unsure whether she is in the situation she foresaw she would be in.

The Reflection Principle must be amended to make sense of the fact that we might now know more about the situation we foresee ourselves being in than we know at the time. Let X summarize everything we know about our status in that foreseen future situation and Y summarize everything else one presently knows.

               New Reflection: Pr(A|pr+ in X) = Pr+(A |Y & X) where pr+ is the statement that Pr( ) is one’s future probability assignment.

What happens to Sleeping Beauty? Well, Beauty only knows that she will awaken on Monday, her present probability in HEADS conditional on the fact that she will come to have a 1/3 probability shouldn’t be 1/3, because in the only situation she foresees herself being in, she will have a ½ probability in HEADS conditional on it being that situation (i.e. MONDAY).

Thus Pr(HEADS|pr+ on Monday) = Pr+(HEADS|MONDAY) = 1/2


One thing that is interesting about this explanation is that it is available only to Thirders. For Halfers of the traditional Lewisian sort, this explanation won’t be available. That is one more problem for Lewisian Halfers. 

No comments:

Post a Comment