I’m pretty much a novice at decision theory, although I’m competent at game theory (and mechanism design), but some of the arguments used to motivate using UDT seem flawed. In particular the “you play prisoner’s dilemma against a copy of yourself” example against CDT seems like its solution relies less on UDT than on the ability to self-modify.
It is true that if you are capable of self-modifying to UDT, you can solve the problem of defecting against yourself by doing so. However if you’re capable of self-modifying, you’re also capable of arbitrarily strong precommitments, which solves the issue without (really) changing decision theories. For example, you can just precommit to “I will cooperate with everyone who shares this precommitment” (for some well-defined “cooperate”*). Then when you’re copied, your copy shares the precommitment and you’re good.
Does that sounds about right or am I missing something?
*regardless of decision theory, you probably wouldn’t want to cooperate with someone who plans to use any resources she obtains to harm you as much as possible, for example.
A CDT agent who is given the choice to self-modify at time t will not self-modify completely into a UDT agent. After self-modification, the agent will one-box in a Newcomb’s problem where Omega made its prediction by examining the agent after time t, and will two-box in a Newcomb’s problem where Omega made its prediction by examining the agent before time t, even if Omega knew that the agent would have the opportunity to self-modify.
In other words, the CDT agent can self-modify to stop updating, but it isn’t motivated to un-update.
You’re capable of arbitrarily strong precommitments.
Using UDT is one way of going about making those precommitments. You precommit to make the decision that you expect will give you the most utility, on average, even if CDT says that you will do worse this time around.
I don’t have one! I’m not brave enough to start coming up with new decision theories while not knowing very much about decision theories. But would I be correct in assuming that this would also mean that the literature definition implies that a CDT agent also can’t choose to become a UDT one? (As that seems to me equivalent to a big precommitment to act as a UDT agent.)
I’m pretty much a novice at decision theory, although I’m competent at game theory (and mechanism design), but some of the arguments used to motivate using UDT seem flawed. In particular the “you play prisoner’s dilemma against a copy of yourself” example against CDT seems like its solution relies less on UDT than on the ability to self-modify.
It is true that if you are capable of self-modifying to UDT, you can solve the problem of defecting against yourself by doing so. However if you’re capable of self-modifying, you’re also capable of arbitrarily strong precommitments, which solves the issue without (really) changing decision theories. For example, you can just precommit to “I will cooperate with everyone who shares this precommitment” (for some well-defined “cooperate”*). Then when you’re copied, your copy shares the precommitment and you’re good.
Does that sounds about right or am I missing something?
*regardless of decision theory, you probably wouldn’t want to cooperate with someone who plans to use any resources she obtains to harm you as much as possible, for example.
A CDT agent who is given the choice to self-modify at time t will not self-modify completely into a UDT agent. After self-modification, the agent will one-box in a Newcomb’s problem where Omega made its prediction by examining the agent after time t, and will two-box in a Newcomb’s problem where Omega made its prediction by examining the agent before time t, even if Omega knew that the agent would have the opportunity to self-modify.
In other words, the CDT agent can self-modify to stop updating, but it isn’t motivated to un-update.
Using UDT is one way of going about making those precommitments. You precommit to make the decision that you expect will give you the most utility, on average, even if CDT says that you will do worse this time around.
The literature largely defines CDT as incapable of precommitments. If you want to propose a specific model of how to choose commitments, just do it.
I don’t have one! I’m not brave enough to start coming up with new decision theories while not knowing very much about decision theories. But would I be correct in assuming that this would also mean that the literature definition implies that a CDT agent also can’t choose to become a UDT one? (As that seems to me equivalent to a big precommitment to act as a UDT agent.)