I’ve come around to the view that global optimisation for a non-trivial objective function in the real world is grossly intractable, so mechanistic utility maximisers are not actually permitted by the laws of physics[1][2].
My remaining uncertainty around expected utility maximisers as a descriptive model of consequentialist systems is whether the kind of hybrid optimisation (mostly learned heuristics, some local/task specific planning/search) that real world agents perform converges towards better approximating argmax wrt (the expected value of) a simple utility function over agent/environment states.
Excluding exotic phenomena like closed timelike curves, the laws of physics of our universe do not seem to permit the construction of computers that can compute optimal actions to maximise non-trivial utility functions over states of the real world within the lifetime of the universe.
Mechanistic Utility Maximisers are Infeasible
I’ve come around to the view that global optimisation for a non-trivial objective function in the real world is grossly intractable, so mechanistic utility maximisers are not actually permitted by the laws of physics[1][2].
My remaining uncertainty around expected utility maximisers as a descriptive model of consequentialist systems is whether the kind of hybrid optimisation (mostly learned heuristics, some local/task specific planning/search) that real world agents perform converges towards better approximating argmax wrt (the expected value of) a simple utility function over agent/environment states.
Excluding exotic phenomena like closed timelike curves, the laws of physics of our universe do not seem to permit the construction of computers that can compute optimal actions to maximise non-trivial utility functions over states of the real world within the lifetime of the universe.
I might be wrong on this, I don’t know physics. I’m mostly relying on intuitions re: combinatorial explosions and exponential complexity.