Do we know in advance which questions are going to be answered? Do we know agent A from your OP isn’t going to end up crying in step 2?
I apologize if I’m just asking about standard CS assumptions here. I have a CS bachelor’s degree but I don’t remember ever discussing programs that can call halting oracles. Either my memory is faulty (more likely) or I chose the wrong electives.
It’s easy to write a universe program U that would make A cry. The post only proves that A won’t end up crying for one specific U, and outlines an argument why it won’t cry in some other similar problems.
The class of universes where A doesn’t end up crying is supposed to roughly correspond to the informal class of “fair” decision problems where the agent’s action is the only thing that determines the resulting utility, but I have no strong argument why it’s a good formalization of that class.
One example of an “unfair” decision problem would be a universe that rewarded you for having a specific algorithm, rather than for returning a specific value. Such “unfair” problems can be made up to punish any decision theory you can come up with, so they’re probably not a good test case.
Is anyone at all working on classes of “unfair” problems, such as ones that give different utilities based upon the amount of time spent computing? Or ones that take into consideration any type of resource used to make that decision (energy or memory). This class seems important to me and less arbitrary than “unfair” problems that punish specific algorithms.
Wei Dai has a tentative decision theory that covers some of those cases. I didn’t find it very convincing, but it’s likely that I overlooked something. Any work on such problems would be very welcome, of course.
I’ll have a think. An optimal decision maker for all scenarios seems impossible if your utility is reduced by an amount proportional to the time take to make the decision (“solving death” has this structure, less people die if you solve it earlier). The best in general I can think of is an infinite table mapping scenarios to a the decision computed by something like your UDT + oracle for that scenario. And this can be beaten in each individual scenario by a specialised algorithm for that scenario, that needs no look up.
And it still has an infinite quantity which I don’t like in my theories that I might want to connect to the real world one day (and requires an infinite amount of precomputation).
I wonder if there is a quality apart from strict optimality that we need to look for. Making the optimal decision in most problems( what is the correct weighting of scenarios)? Making the right decision eventually?
Anyway I’ll think some more. It is definitely thornier and nastier than “fair” problems.
Do we know in advance which questions are going to be answered? Do we know agent A from your OP isn’t going to end up crying in step 2?
I apologize if I’m just asking about standard CS assumptions here. I have a CS bachelor’s degree but I don’t remember ever discussing programs that can call halting oracles. Either my memory is faulty (more likely) or I chose the wrong electives.
It’s easy to write a universe program U that would make A cry. The post only proves that A won’t end up crying for one specific U, and outlines an argument why it won’t cry in some other similar problems.
The class of universes where A doesn’t end up crying is supposed to roughly correspond to the informal class of “fair” decision problems where the agent’s action is the only thing that determines the resulting utility, but I have no strong argument why it’s a good formalization of that class.
One example of an “unfair” decision problem would be a universe that rewarded you for having a specific algorithm, rather than for returning a specific value. Such “unfair” problems can be made up to punish any decision theory you can come up with, so they’re probably not a good test case.
Thanks for explaining.
Is anyone at all working on classes of “unfair” problems, such as ones that give different utilities based upon the amount of time spent computing? Or ones that take into consideration any type of resource used to make that decision (energy or memory). This class seems important to me and less arbitrary than “unfair” problems that punish specific algorithms.
Wei Dai has a tentative decision theory that covers some of those cases. I didn’t find it very convincing, but it’s likely that I overlooked something. Any work on such problems would be very welcome, of course.
I’ll have a think. An optimal decision maker for all scenarios seems impossible if your utility is reduced by an amount proportional to the time take to make the decision (“solving death” has this structure, less people die if you solve it earlier). The best in general I can think of is an infinite table mapping scenarios to a the decision computed by something like your UDT + oracle for that scenario. And this can be beaten in each individual scenario by a specialised algorithm for that scenario, that needs no look up.
And it still has an infinite quantity which I don’t like in my theories that I might want to connect to the real world one day (and requires an infinite amount of precomputation).
I wonder if there is a quality apart from strict optimality that we need to look for. Making the optimal decision in most problems( what is the correct weighting of scenarios)? Making the right decision eventually?
Anyway I’ll think some more. It is definitely thornier and nastier than “fair” problems.
I recently made some progress on your question. Section 4 seems to be the most relevant.