= 27d567bc2d48d9f40e9ccc20ea370b15)
Will_Sawin comments on Rationality Quotes: June 2011 - Less Wrong
You are viewing a comment permalink. View the original post to see all comments and the full post content.
= 783df68a0f980790206b9ea87794c5b6)
Loading…= b715d02e85f7b3b4ae65ca3d3e450868)
Subscribe to RSS Feed
Powered by Reddit
= f037147d6e6c911a85753b9abdedda8d)
You are viewing a comment permalink. View the original post to see all comments and the full post content.
Comments (470)
Perhaps an example of what I mean will be helpful.
Suppose your friend is kidnapped and being held for ransom. Naive consequentialism says you should pay because you value his life more then the money. TDT says you shouldn't pay because paying counterfactually causes him to be kidnapped.
Note how in the scenario the TDT argument sounds very deontological.
It sounds deontological, but it isn't. It's consequentialist. It evaluates options according to their consequences.
"Consequences" only in a counterfactual world. I don't see how you can call this consequentialist without streching the term to the point that it could include nearly any morality system. In particular by your definition Kant's categorical imperative is consequentialist since it involves looking at the consequences of your actions in the hypothetical world where everyone performs them.
Yes, in that TDT-like decision/ethical theories are basically "consequentialism in which you must consider 'acausal consequences'".
While it may seem strange to regard ethical theories that apply Kant's CI as "consequentialist", it's even stranger to call them deontological, because there is no deontic-like "rule set" they can be said to following; it's all simple maximization, albeit with a different definition of what you count as a benefit. TDT, for example, considers not only what your action causes (in the technical sense of future results), but the implications of the decision theory you instantiate having a particular output.
(I know there are a lot of comments I need to reply to, I will get to them, be patient.)
It certainly is strange even if it is trivially possible. Any 'consequentialist' system can be implemented in a singleton deontological 'rule set'. In fact, that's the primary redeeming feature of deontology. Kind of like the best thing about Java is that you can use it to implement JRuby and bypass all of Java's petty restrictions and short sighted rigidly enforced norms.
Both CDT and TDT compare counter-factuals, they just take their counter-factual from different points in the causal graph.
In both cases, while computing them you never assume anything which you know to be false, whereas Kant is not like that. (Just realised, I'm not sure this is right).
Counterfactual mugging and the ransom problem I mentioned in the great-grandparent are both cases where TDT requires you to consider consequences of counterfactuals you know didn't happen. Omega's coin didn't come up heads, and your friend has been kidnapped. Nevertheless you need to consider the consequences of your policy in those counterfactual situations.
I think counterfactual mugging was originally brought up in the context of problems which TDT doesn't solve, that is it gives the obvious but non-optimal answer. The reason is that regardless of my counterfactual decision Omega still flips the same outcome and still doesn't pay.
There are two rather different things both going under the name counterfactuals.
One is when I think of what the world would be like if I did something that I'm not going to do.
Another is when I think of what the world would be like if something not under my control had happened differently, and how my actions affect that.
They're almost orthogonal, so I question the utility of using the same word.
Well, I've been consistently using the word "conterfactual" in your second sense.
Well that might explain some of our miscommunication. I'll go back and check.
This makes sense using the first definition, at least, according to TDT it does.
This is clearly using the first definition.
This only makes sense with the second, and should probably be UDT rather than TDT - the original TDT didn't get the right answer on the counterfactual mugging.
Sorry, I meant something closer to UDT.
Alright cool. So I think that's what's going on - we all agree but were using different definitions of counterfactuals.
You need a proof-system to ensure that you never assume anything which you know to be false.
ADT and some related theories have achieved this. I don't think TDT has.
What I meant by that statement was the idea that CDT works by basing counterfactuals on your action, which seems a reasonable basis for counterfactuals since prior to making your decision you obviously don't know what your action will be. TDT similarly works by basing counterfactuals on your decision, which you also don't know prior to making it.
Kant, on the other hand, bases his counter-factuals on what would happen if everyone did that, and it is possible that his will involve assuming things I know to be false in a sense that CDT and TDT don't (e.g. when deciding whether to lie I evaluate possible worlds in which everyone lies and in which everyone tells the truth, both of which I know not to be the case).
Well here is the issue.
Let's say I have to decide what to do at 2'o'clock tomorrow. If I light a stick of dynamite, I will be exploded. If I don't, then I won't. I can predict that I will, in fact, not light a stick of dynamite tomorrow. I will then know that one of my counterfactuals is true and one is false.
This can mess up the logic of decision-making. There are . This ensures that you can never figure out a decision before making it, which makes things simpler.
I'm not sure if this contradicts what you've said.
And I would agree exactly with your analysis about what's wrong with Kant, and how that's different from CDT and TDT.
I'm not sure I agree with myself. I think my analysis makes sense for the way TDT handles Newcomb's problem or Prisoner's dilemma, but it breaks down for Transparent Newcomb or Parfit's Hitch-hiker. In those cases, owing to the assistance of a predictor, it seems like it is actually possible to know your decision in advance of making it.
Well you always know that one of your counterfactuals is true.
and Transparent Newcomb is a bit weird because one of the four possible strategies just explodes it.
There is no need to make that assumption. The whole collection of possible decisions could be located on an impossible counterfactual. Incidentally, this is one way of making sense of Transparent Newcomb.
You could possibly fix that by saying Omega isn't perfect, but his predictions are correlated enough with your decision to make precomittment possible.
Yes. However that decision theory is wrong and dumb so we can ignore it. In particular, it never produces factuals, only counterfactuals.