I saw this post from EY a while ago and felt kind of repulsed by it:
I no longer feel much of a need to engage with the hypothesis that rational agents mutually defect in the oneshot or iterated PD. Perhaps you meant to analyze causal-decision-theory agents?
Never mind the factual shortcomings, I’m mostly interested in the rejection of CDT as rational. I’ve been away from LW for a while and wasn’t keeping up on the currently popular beliefs on this site, and I’m considering learning a bit more about TDT (or UDT or whatever the current iteration is called). I have a feeling this might be a huge waste of time though, so before I dive into the subject I would like to confirm that TDT has objectively been proven to be clearly superior to CDT, by which I (intuitively) mean:
There exist no problems shown to be possible in real life for which CDT yields superior results.
There exists at least one problem shown to be possible in real life for which TDT yields superior results.
“Shown to be possible in real life” excludes Omega, many-worlds, or anything of similar dubiousness. So has this been proven? Also, is there any kind of reaction from the scientific community in regards to TDT/UDT?
The question “which decision theory is superior?” has this flavor of “can my dad beat up your dad?”
CDT is what you use when you want to make decisions from observational data or RCTs (in medicine, and so on).
TDT is what you use when “for some reason” your decisions are linked to what counterfactual versions/copies of yourself decided. Standard CDT doesn’t deal with this problem, because it lacks the language/notation to talk about these issues. I argue this is similar to how EDT doesn’t handle confounding properly because it lacks the language to describe what confounding even means. (Although I know a few people who prefer a decision algorithm that is in all respects isomophic to CDT, but which they prefer to call EDT for I guess reasons having to do with the formal epistemology they adopted. To me, this is a powerful argument for not adopting a formal epistemology too quickly :) )
I think it’s more fruitful to think about the zoo of decision theories out there in terms of what they handle and what they break on, rather than in terms of anointing some of them with the label “rational” and others with the label “irrational.” These labels carry no information. There is probably no total ordering from “best to worst” (for example people claim EDT correctly one boxes on Newcomb, whereas CDT does not. This does not prevent EDT from being generally terrible on the kinds of problems CDT handles with ease due to a worked out theory of causal inference).
I don’t like the notion of using different decision theories depending on the situation, because the very idea of a decision theory is that it is consistent and comprehensive. Now if TDT were formulated as a plugin that seamlessly integrated into CDT in such a way that the resulting decision theory could be applied to any and all problems and would always yield optimal results, then that would be reason for me to learn about TDT. However, from what I gathered this doesn’t seem to be the case?
TDT performs exactly as well as CDT on the class of problems CDT can deal with, because for those problems it essentially is CDT. So in practice you just use normal CDT algorithms except for when counterfactual copies of yourself are involved. Which is what TDT does.
I argue that there’s an mapping in the opposite direction: if you add extra nodes to any problem that looks like a problem where TDT and CDT disagree, and adjust which node is the decision node, then you can make CDT and TDT agree (and CDT give the “TDT solution”). This is obvious in the case of Newcomb’s Problem, for example.
I guess it’s true that CDT needed lots of ideas to work. TDT has one idea: “link counterfactual decisions together.” So it is not an unreasonable view that TDT is an addendum to CDT, and not vice versa, since CDT is intellectually richer.
This is essentially what the TDT paper argues. It’s been a while since I’ve read it, but at the time I remember being sufficiently convinced that it was strictly superior to both CDT and EDT in the class of problems that those theories work with, including problems that reflect real life.
I think people have slightly misunderstood what I was referring to with this:
There exist no problems shown to be possible in real life for which CDT yields superior results.
There exists at least one problem shown to be possible in real life for which TDT yields superior results.
My question was whether there is a conclusive, formal proof for this, not whether this is widely accepted on this site (I already realized TDT is popular). If someone thinks such a proof is given somewhere in an article (this one?) then please direct me to the point in the article where I can find that proof. I’m very suspicious about this though, since the wiki makes blatantly false claims, e.g. that TDT performs better in one-shot PD than CDT, while in fact it can only perform better if access to source coude is given. So the wiki article feels more like promotion than anything.
Also, I would be very interested to hear about what kind of reaction from the scientific community TDT has received. Like, very very interested.
Parfit’s hitchhiker looks like a thinly veiled Omega problem to me. At the very least, considering the lack of scientific rigorousness in Ekman’s research, it should count as quite dubious, so adopting a new decision theory on the basis of that particular problem does not seem rational to me.
Yes, it’s a Newcomb-like problem. Anything where one agent predicts another is. People predict other people, with varying degrees of success, in the real world. Ignoring that when looking at decision theories seems silly to me.
I hold the belief that Newcomb, regardless of Omega’s accuracy, is impossible in the universe I currently live in. Also, this is not what this discussion is about, so please refrain from derailing it further.
Newcomb-like problems are the ones where TDT outperforms CDT. If you consider these problems to be impossible, and won’t change your mind, then you can’t believe that TDT satisfies your requirements.
I saw this post from EY a while ago and felt kind of repulsed by it:
Never mind the factual shortcomings, I’m mostly interested in the rejection of CDT as rational. I’ve been away from LW for a while and wasn’t keeping up on the currently popular beliefs on this site, and I’m considering learning a bit more about TDT (or UDT or whatever the current iteration is called). I have a feeling this might be a huge waste of time though, so before I dive into the subject I would like to confirm that TDT has objectively been proven to be clearly superior to CDT, by which I (intuitively) mean:
There exist no problems shown to be possible in real life for which CDT yields superior results.
There exists at least one problem shown to be possible in real life for which TDT yields superior results.
“Shown to be possible in real life” excludes Omega, many-worlds, or anything of similar dubiousness. So has this been proven? Also, is there any kind of reaction from the scientific community in regards to TDT/UDT?
The question “which decision theory is superior?” has this flavor of “can my dad beat up your dad?”
CDT is what you use when you want to make decisions from observational data or RCTs (in medicine, and so on).
TDT is what you use when “for some reason” your decisions are linked to what counterfactual versions/copies of yourself decided. Standard CDT doesn’t deal with this problem, because it lacks the language/notation to talk about these issues. I argue this is similar to how EDT doesn’t handle confounding properly because it lacks the language to describe what confounding even means. (Although I know a few people who prefer a decision algorithm that is in all respects isomophic to CDT, but which they prefer to call EDT for I guess reasons having to do with the formal epistemology they adopted. To me, this is a powerful argument for not adopting a formal epistemology too quickly :) )
I think it’s more fruitful to think about the zoo of decision theories out there in terms of what they handle and what they break on, rather than in terms of anointing some of them with the label “rational” and others with the label “irrational.” These labels carry no information. There is probably no total ordering from “best to worst” (for example people claim EDT correctly one boxes on Newcomb, whereas CDT does not. This does not prevent EDT from being generally terrible on the kinds of problems CDT handles with ease due to a worked out theory of causal inference).
I don’t like the notion of using different decision theories depending on the situation, because the very idea of a decision theory is that it is consistent and comprehensive. Now if TDT were formulated as a plugin that seamlessly integrated into CDT in such a way that the resulting decision theory could be applied to any and all problems and would always yield optimal results, then that would be reason for me to learn about TDT. However, from what I gathered this doesn’t seem to be the case?
TDT performs exactly as well as CDT on the class of problems CDT can deal with, because for those problems it essentially is CDT. So in practice you just use normal CDT algorithms except for when counterfactual copies of yourself are involved. Which is what TDT does.
I argue that there’s an mapping in the opposite direction: if you add extra nodes to any problem that looks like a problem where TDT and CDT disagree, and adjust which node is the decision node, then you can make CDT and TDT agree (and CDT give the “TDT solution”). This is obvious in the case of Newcomb’s Problem, for example.
I guess it’s true that CDT needed lots of ideas to work. TDT has one idea: “link counterfactual decisions together.” So it is not an unreasonable view that TDT is an addendum to CDT, and not vice versa, since CDT is intellectually richer.
This is essentially what the TDT paper argues. It’s been a while since I’ve read it, but at the time I remember being sufficiently convinced that it was strictly superior to both CDT and EDT in the class of problems that those theories work with, including problems that reflect real life.
I think people have slightly misunderstood what I was referring to with this:
My question was whether there is a conclusive, formal proof for this, not whether this is widely accepted on this site (I already realized TDT is popular). If someone thinks such a proof is given somewhere in an article (this one?) then please direct me to the point in the article where I can find that proof. I’m very suspicious about this though, since the wiki makes blatantly false claims, e.g. that TDT performs better in one-shot PD than CDT, while in fact it can only perform better if access to source coude is given. So the wiki article feels more like promotion than anything.
Also, I would be very interested to hear about what kind of reaction from the scientific community TDT has received. Like, very very interested.
Then no. In “normal” situations CDT does as well as anything else.
Didn’t the paper show TDT performing better than CDT in Parfit’s Hitchhiker?
That might count as being of similar dubiousness, although I like this quote by Eliezer arguing otherwise:
Parfit’s hitchhiker looks like a thinly veiled Omega problem to me. At the very least, considering the lack of scientific rigorousness in Ekman’s research, it should count as quite dubious, so adopting a new decision theory on the basis of that particular problem does not seem rational to me.
Yes, it’s a Newcomb-like problem. Anything where one agent predicts another is. People predict other people, with varying degrees of success, in the real world. Ignoring that when looking at decision theories seems silly to me.
What do you do in Newcomb’s problem if Omega has a 45% chance of mispredicting you?
Algebra.
I’d start calling myself Omega Prime and making the reverse prediction to just say I’m smarter than Omega.
You’d then have a 55% chance of mispredicting (slightly worse than chance, where the 45% Omega is slightly better than chance).
Looks like I’d first have to start reading what people write correctly!
I hold the belief that Newcomb, regardless of Omega’s accuracy, is impossible in the universe I currently live in. Also, this is not what this discussion is about, so please refrain from derailing it further.
It’s highly relevant to your second point.
Newcomb-like problems are the ones where TDT outperforms CDT. If you consider these problems to be impossible, and won’t change your mind, then you can’t believe that TDT satisfies your requirements.