I think I have a fairly straightforward theory of logical counterfactuals, which I believe makes something very much like PCDT work (except instead of a C(A|B) construct, it’s a C(A|B=C) construct; but I assume this is just an irrelevant syntactic change). Would it be worth a writeup? Rough estimates of its properties (haven’t fully formalized yet so it might be up for revision):
It avoids the dutch-book of CDT by keeping its conditionals in line with its decision counterfactuals (insofar as its conditionals are even meaningful; conditionals over actions are hard to define due to having knowledge over your actions—but e.g. if we add epsilon-exploration and don’t have any shenanigans that separate the epsilon-exploration actions from the deliberate actions, it’d end up equivalent).
It handle troll’s bridge correctly by reasoning counterfactually.
It one-boxes in Newcomb’s problem and cooperates with itself in prisoners dilemma (while of course defecting against cooperatebot and defectbot).
It can do a proof-based mode, but it can also do a model-based mode that in the case of full knowledge is provably equivalent in outputs, and can be efficiently computed for toy problems.
It doesn’t have any learning or tiling theory.
Anyway I thought I’d ask before I write it up, since my solution is fairly basic, so I wouldn’t be surprised if someone had derived it already since your post.
Actually, whether it one-boxes or two-boxes in Newcomb’s problem depends on the setup. How is Newcomb’s problem usually formalized in these logic-based settings?
What’s the status on this?
I think I have a fairly straightforward theory of logical counterfactuals, which I believe makes something very much like PCDT work (except instead of a C(A|B) construct, it’s a C(A|B=C) construct; but I assume this is just an irrelevant syntactic change). Would it be worth a writeup? Rough estimates of its properties (haven’t fully formalized yet so it might be up for revision):
It avoids the dutch-book of CDT by keeping its conditionals in line with its decision counterfactuals (insofar as its conditionals are even meaningful; conditionals over actions are hard to define due to having knowledge over your actions—but e.g. if we add epsilon-exploration and don’t have any shenanigans that separate the epsilon-exploration actions from the deliberate actions, it’d end up equivalent).
It handle troll’s bridge correctly by reasoning counterfactually.
It one-boxes in Newcomb’s problem and cooperates with itself in prisoners dilemma (while of course defecting against cooperatebot and defectbot).
It can do a proof-based mode, but it can also do a model-based mode that in the case of full knowledge is provably equivalent in outputs, and can be efficiently computed for toy problems.
It doesn’t have any learning or tiling theory.
Anyway I thought I’d ask before I write it up, since my solution is fairly basic, so I wouldn’t be surprised if someone had derived it already since your post.
Actually, whether it one-boxes or two-boxes in Newcomb’s problem depends on the setup. How is Newcomb’s problem usually formalized in these logic-based settings?
Or I guess it would be called RCDT, since I’m proposing a specific class of counterfactuals? I’m not sure I understand the PCDT/RCDT distinction.