Context: Causal Decision Theory (CDT) = Choose the action that cause the best expected outcome Evidential Decision Theory (EDT) = Choose the action that is the best evidence of best expected outcome
In Does Dissolving Newcomb’s Paradox Matter I wrote a bit about decision theory. Specifically I mentioned the smoking lesion case as one where evidential decision theory fails because it gives the wrong answer. On reflection, I’m not sure that’s correct and I should probably start using a different example when trying to show the flaws of EDT.
Why does the smoking lesion case not really work? The case is set up roughly like this
there’s a specific genetic lesion someone can have
there’s no way for a person to test if they have the lesion
that lesion causes them to enjoy smoking
the lesion also causes cancer
the act of smoking itself is otherwise not linked to cancer
The correct action for an agent here is to smoke if they have the lesion. Doing so won’t change their cancer risk. It will give them some pleasure. Hence it’s net positive. CDT suggests the correct action. Smoking gives you pleasure. Smoking does not cause an increase in cancer risk because you either already have the lesion or not. Hence you should smoke. I was taught that EDT gives the incorrect answer, that smoking is correlated with cancer and so you shouldn’t do it regardless of the causal chain. This actually doesn’t feel right. The thing is it’s not the act of smoking that correlates with cancer. It’s the fact that you enjoy smoking. Hence the correlation is already there regardless of whether or not you choose to smoke. It’s the fact that you would enjoy smoking that correlates. Hence EDT probably also correctly recommends you smoke. It turns out this is called the Tickle Defense of EDT. I don’t think I buy the whole thing, especially not for other cases like Newcombs (Omega is predicting your action, not your inclination), but it dpes seem to work here.
This could be different if you tweak the setup a bit. If you can’t perfectly introspect, then smoking does give you evidence and the example works again but that seems a bit inelegant. Still works, but probably not the example I want to use when introducing the concept to people. I think in the future I’ll stick with different examples. For showing the flaws in CDT I typically use “playing a prisoners dilemma against a copy of yourself, but the copy already made their decision” or Newcomb’s Problem. For EDT I’ll need to find something else. Ideally something that both doesn’t rely on prediction of the agent’s action and in doing so trigger various free-will questions and doesn’t smuggle in self-knowledge assumptions.
My response to the smoking lesion question is:
If the lesion doesn’t affect your ability to reason, then deciding to smoke as the result of a reasoning process isn’t correlated with cancer, even though deciding to smoke in general is, so you should smoke.
If the lesion does affect your ability to reason, the question is ill defined because it may not be possible for you to execute some strategies, or only possible to execute them with some probability.
Of course “as the result of a reasoning process” needs to be phrased more precisely (every decision can be said to be the result of some reasoning process). Maybe “when you make some particular deduction, compared to a counterfactual world where you did not make that deduction”.
Can’t they just check whether they enjoy smoking?
In some versions or discussions of the Smoking Lesion problem, yes. In others, no. There does not appear to be consensus on what the actual scenario is.
The Tickle Defense arguably makes EDT behave like CDT. The topic EDT=CDT was discussed in this forum before. Though that was years ago, since decision theory has somewhat fallen out of fashion.
Typo, should be “does”