3 ms·

I mean, assuming the behavior satisfies appropriate coherence conditions... (which actual human behavior doesn't as Misdicorl points out, but that's another mat

I mean, assuming the behavior satisfies appropriate coherence conditions... (which actual human behavior doesn't as Misdicorl points out, but that's another matter).

a1369209993 8y ago
No, any behavior can be mapped to "I prefer world-histories where I do A at time 1, B at time 2, C at time 3, ...". Your coherence conditions are the constraints on the notion of preferences without which trivial-behaviorist preferences can be constructed. Edit: but yes, human behavior (often) doesn't satisfy von Neumann-Morgenstern utility axioms (the coherence conditions for scale-free origin-free utility functions as prefered by behavioral economics).
- eli_gottlieb 8y ago
  >Edit: but yes, human behavior (often) doesn't satisfy von Neumann-Morgenstern utility axioms (the coherence conditions for scale-free origin-free utility functions as prefered by behavioral economics). In specific, the VNM axioms assume that preferences aren't over histories at all, but instead over states, which severely constrains the kind of outcome-generating process they allow an agent to care about. They effectively state, "only the end-point matters, not the trajectory". This doesn't really make sense for humans in the first place, since we're made of meat and need to do things like eat and sleep cyclically. What it does make sense for is bank accounts, which is why economists think it makes sense.
  - a1369209993 8y ago
    > the VNM axioms assume that preferences aren't over histories at all, but instead over states That depends on how you formulate the axioms; or, equivalently, how you formulate the state. Eg, you can construct a vNM agent that strongly avoids "outcomes" where it spend a week awake, even if the end state of the world the agent is operating in retains no information about that. The "outcomes" A_i of the vNM axioms are (potentially) the entire observable history of the agent's universe.
- Sniffnoy 8y ago
  Such a preference ordering has nothing to do with real numbers or expected values, though. Orderings don't by themselves lead to real numbers without some sort of countability or Archimedean requirement, and they certainly don't lead to expected values without quite a bit more. (IMO Savage's Theorem[0] provides a better foundation than VNM, but that's nitpicking.) [0]See my article on it, if you're unfamiliar: https://www.lesswrong.com/posts/5J34FAKyEmqKaT7jt/a-summary-of-savage-s-foundations-for-probability-and https://www.lesswrong.com/posts/5J34FAKyEmqKaT7jt/a-summary-...

I mean, assuming the behavior satisfies appropriate coherence conditions... (which actual human behavior doesn't as Misdicorl points out, but that's another matter).

a1369209993 8y ago

No, any behavior can be mapped to "I prefer world-histories where I do A at time 1, B at time 2, C at time 3, ...". Your coherence conditions are the constraints on the notion of preferences without which trivial-behaviorist preferences can be constructed. Edit: but yes, human behavior (often) doesn't satisfy von Neumann-Morgenstern utility axioms (the coherence conditions for scale-free origin-free utility functions as prefered by behavioral economics).

eli_gottlieb 8y ago
>Edit: but yes, human behavior (often) doesn't satisfy von Neumann-Morgenstern utility axioms (the coherence conditions for scale-free origin-free utility functions as prefered by behavioral economics). In specific, the VNM axioms assume that preferences aren't over histories at all, but instead over states, which severely constrains the kind of outcome-generating process they allow an agent to care about. They effectively state, "only the end-point matters, not the trajectory". This doesn't really make sense for humans in the first place, since we're made of meat and need to do things like eat and sleep cyclically. What it does make sense for is bank accounts, which is why economists think it makes sense.
- a1369209993 8y ago
  > the VNM axioms assume that preferences aren't over histories at all, but instead over states That depends on how you formulate the axioms; or, equivalently, how you formulate the state. Eg, you can construct a vNM agent that strongly avoids "outcomes" where it spend a week awake, even if the end state of the world the agent is operating in retains no information about that. The "outcomes" A_i of the vNM axioms are (potentially) the entire observable history of the agent's universe.
Sniffnoy 8y ago
Such a preference ordering has nothing to do with real numbers or expected values, though. Orderings don't by themselves lead to real numbers without some sort of countability or Archimedean requirement, and they certainly don't lead to expected values without quite a bit more. (IMO Savage's Theorem[0] provides a better foundation than VNM, but that's nitpicking.) [0]See my article on it, if you're unfamiliar: https://www.lesswrong.com/posts/5J34FAKyEmqKaT7jt/a-summary-of-savage-s-foundations-for-probability-and https://www.lesswrong.com/posts/5J34FAKyEmqKaT7jt/a-summary-...