Any thoughts? @wiglet1981.bsky.social @bweatherson.bsky.social @larabuchak.net @erikangner.com @cailinmeister.bsky.social @jweisber.bsky.social @harveylederman.bsky.social @lastpositivist.bsky.social There must be an old discussion of this by Jeffrey somewhere...
"Don't care at all?" Economists have been discussing the issue since `Rules rather than discretion: the inconsistency of optimal plans' Kydland & Prescott J. Pol. Econ. 1977, cited 13000 times. The link to Newcomb was noted in 1982. Broome informed philosophers in 1989.
Of course economists care about making decisions when someone else’s action is based on a prediction of what you will do. But that’s not the question I was asking. I’m wondering about models that make probabilities of states depend on actions vs models that require states to encode causal facts.
explanations in this thread all seem close enough for the comments section so far... what do you mean by "states encode causal facts"?
Consider a student studying for a test. If you write the states as pass and fail, studying will be dominated. You need a state in which the student passes if and only if she passes.
Philosophers know this too, of course, but they respond by considering probabilities of states (still pass or fail) that depend on the action.
hmm. i'm not sure i follow your meaning yet. it seems interesting though and i'd like to read more about it
The philosophy literature is around a contrast between what they call evidential and causal decision theory, where Newcomb’s paradox is a central example. In Gilboa’s decision theory textbook, he has a chapter called something like what is a state? that’s all about handling the issue in ala savage
My knowledge is out of date, but there was a strand of philosophy literature that tried to capture causal decision theory in Savage's framework but concluded you can't. But the approach they considered is different from the one I think you are proposing.
Causal decision theorists usually replace talk about states with some kind of dependency relationship between acts and outcomes. Are you asking: "Why not call such a dependency relation a state?" (I don't have Gilboa's book to hand.)
Yes, that’s the thought. My hunch is that, whatever turns out to be the right model theory for the conditionals can be turned into a Savage state. (This is based on no serious work ofc).
The relation between the two literatures, as far as I can tell, is that Savage done according to the modeling principles Savage and Gilbo outline is a causal decision theory, and evidential decision theory is undreamed of by our tribe. My question was, are we missing something?
Wow, talk about being divided by a common language. My guess is that almost all decision theory-ish philosophers think economists are all evidential decision theorists with no interest in causal decision theory
Huh, I’d always thought philosophers took economists to all be causal theorists since (a) Newcomb is a variant of Prisoners’ Dilemma, and (b) economists say to defect in (one shot) Prisoners’ Dilemma.
That was my worry!
In case you’re not following because I’m being too informal: if I were modeling the test problem, I’d write down four states PP, PF, FP, and FF (first letter is consequence of if I study, second if I don’t). Then the second and third states capture the possibility my action affects the outcome.
nah think something else is confusing me, thanks though 🙏
is that "passes if and only if she studies"? sorry for typo-picking but this is something where I'm far enough out to sea that I'm genuinely not sure
Yes, sorry for the typo
yes, not a statement about Scott's exams!!!
In an economists’s version of a game theoretic model, a player’s state space is the set of (exogenous) states of nature cross the strategy spaces of the other players. So the dependence is captured without making the realized state depend on the actual choice.
I’m a control theorist (so, basically, combining worst aspects of economists and philosophers), but here’s how I see it. You want the mapping from acts to consequences to be well-defined (no vicious circles) and measurable (can compute expected utilities).
In usual setups (sequential games with nested information sets, Markov decision processes) you can always do this (acts are functions of state histories).
In nonclassical situations, you can often find a virtual state that allows this to go through. See eg this paper by Ho, Kastner, and Wong: people.eecs.berkeley.edu/~wong/wong_p...
(*) Control theorists are pragmatists philosophically and neoclassicists economically.
I don't know much about econ but my guess is basically the same as Liam's: for philosophers, a theory like decision theory is offered as the True and Final answer to a question (e.g. "What is Rationality?"). So if it gets even one case wrong, it's a failure: not True, not Final. Whereas...
for economists it's just a model to be used in answering further downstream questions. So errant edge-cases aren't so concerning, as long as the model is close enough for government work.
This point about the goals of modeling in the two disciplines is surely correct. But it still leaves my question for innocuous problems like studying for a test: why make the probabilities depend on actions rather than change the state space?
I’m looking for someone to talk me down from my hunch that it’s all a conspiracy by evidnential decision theorists.
Don't we do both though? Best response as a solution concept uses Pr of the actions of other players and nature, correlated equilibrium assigns Pr directly to the terminal states, iterated rationalization lets us update w/ new info. Econs are interested in the constraint of info as it shapes choice.
Correlated equilibrium is an interesting example here, since it does involve thinking about the probability of others’s actions conditional on what I’m supposed to do. But that doesn’t take the form of supposing that conditional distribution changes if I deviate, which is the bit I’m focused on.
Joyce argues that we need a partition-invariant theory like Jeffrey's, and unlike Savage's. I think the idea is, if you have to get the state-space right before you can apply the theory, then... www.jstor.org/stable/188653
philosophically speaking, your theory isn't complete. You still have to say what makes the state-space "right". And this requires bringing in extra resources, like probabilistic independence, that we were trying to derive *from* the theory, rather than build into its foundations.
I can’t quite tell if there are more divided-by-a-common language problems, or if I think that’s a misreading of Savage. He’s clear that his theory is for small worlds, where a small world is a problem that can be treated in isolation with a comprehensive state space.
Everyone agrees that that makes the theory incomplete, since it doesn’t tell you which problems are small worlds.
This also suggests a diagnosis of the difference between the fields: is subjective expected utility supposed to be a fully general theory, or it just the fragment we understand now? My sense is that most Savage-ians take the second option.
Philosophers want a fully general and complete theory, and many have looked to SEU theory for this. We're all clear that Savage's formulation isn't fully general/complete, which is why folks like Jeffrey and Joyce develop further versions meant to bring us closer to the holy grail
What are the non-SEU options in that hunt for a fully general and complete theory?
Saying a bit more: I think philosophers are often concerned to allay suspicions that a concept (e.g. Rationality) is spooky, or supernatural, or irredeemably imprecise by showing that it can be accounted for in respectably naturalistic/precise terms. As long as counterexamples remain...
so will the suspicions, as that may be where the supernatural bodies are buried.
Hearing elsewhere on the thread that actually economists do totally care about it, which leads to the new question of why they do given that as you say they have no reason to
Eh, close enough
Indeed; this morning has been a real emotional rollercoaster bsky.app/profile/mike...
Now hearing that actually they don't care about it after all
lmao