My take on EDT is that it’s, at its core, vague about probability estimation. If the probabilities are accurate forecasts based on detailed causal models of the world, then it works at least as well as CDT. But if there’s even a small gap between the model and reality, it can behave badly.
E.g. if you like vanilla ice cream but the people who get chocolate really enjoy it, you might not endorse an EDT algorithm that thinks of of probabilities as frequency within a reference class. I see the smoking lesion as a more sophisticated version of this same issue.
But then if probabilities are estimated via causal model, EDT has exactly the same problem with Newcomb’s Omega as CDT, because the problem with Omega lies in the incorrect estimation of probabilities when someone can read your source code.
So I see these as two different problems with two different methods of assigning probabilities in an underspecified EDT. This means that I predict there’s an even more interesting version of your example where both methods fail. The causal modelers assume that the past can’t predict their choice, and the reference class forecasters get sidetracked by options that put them in a good reference class without having causal impact on what they care about.
It is not the responsibility of a decision theory to tell you how to form opinions about the world; it should tell you how to use the opinions you have. EDT does not mean reference class forecasting; it means expecting utility according to the opinions you would actually have if you did the thing, not ignoring the fact that doing the thing would give you information.
Or in other words, it means acting on your honest opinion of what will give you the best result, and not a dishonest opinion formed by pretending that your opinions wouldn’t change if you did something.
I think this deflationary conception of decision theory has serious problems. First is that because it doesn’t pin down a decision-making algorithm, it’s hard to talk about what choices it makes—you can argue for choices but you can’t demonstrate them without showing how they’re generated in full. Second is that it introduces more opportunities to fool yourself with verbal reasoning. Third because historically I think it’s resulted in a lot of wasted words in philosophy journals, although maybe this is just objection one again.
My take on EDT is that it’s, at its core, vague about probability estimation. If the probabilities are accurate forecasts based on detailed causal models of the world, then it works at least as well as CDT. But if there’s even a small gap between the model and reality, it can behave badly.
E.g. if you like vanilla ice cream but the people who get chocolate really enjoy it, you might not endorse an EDT algorithm that thinks of of probabilities as frequency within a reference class. I see the smoking lesion as a more sophisticated version of this same issue.
But then if probabilities are estimated via causal model, EDT has exactly the same problem with Newcomb’s Omega as CDT, because the problem with Omega lies in the incorrect estimation of probabilities when someone can read your source code.
So I see these as two different problems with two different methods of assigning probabilities in an underspecified EDT. This means that I predict there’s an even more interesting version of your example where both methods fail. The causal modelers assume that the past can’t predict their choice, and the reference class forecasters get sidetracked by options that put them in a good reference class without having causal impact on what they care about.
It is not the responsibility of a decision theory to tell you how to form opinions about the world; it should tell you how to use the opinions you have. EDT does not mean reference class forecasting; it means expecting utility according to the opinions you would actually have if you did the thing, not ignoring the fact that doing the thing would give you information.
Or in other words, it means acting on your honest opinion of what will give you the best result, and not a dishonest opinion formed by pretending that your opinions wouldn’t change if you did something.
I think this deflationary conception of decision theory has serious problems. First is that because it doesn’t pin down a decision-making algorithm, it’s hard to talk about what choices it makes—you can argue for choices but you can’t demonstrate them without showing how they’re generated in full. Second is that it introduces more opportunities to fool yourself with verbal reasoning. Third because historically I think it’s resulted in a lot of wasted words in philosophy journals, although maybe this is just objection one again.