Your optimizer, whether Bayesian or not, needs to be able to recognize a low point when it hits one, or else it can’t optimize at all! If every point looks the same… (It may learn more about high points, but it must still learn about low points.)
(It may learn more about high points, but it must still learn about low points.)
That’s not how Bayesian optimization works. Broadly, the idea is that we use Bayesian optimization when both calculating the value of the target function at a point and calculating its gradient are both expensive or infeasible. Thus, we instead choose points at which to sample the target function, and the samples train a Gaussian process model (or other nonparametric model of functions) that tells us what the function’s surface looks like. In such a procedure, we obtain the best performance by sampling points where either the expected function value or the model’s variance is particularly high. Thus, we choose points that we know are good, or points where we’re very uncertain, but we never specifically search for low points. We’ll probably encounter some when sampling points of great uncertainty, but we didn’t specifically seek them out.
The model I use to derive that involves looking at lots of dying people who don’t want to die. If we had lots of people lying around saying “I wish I could die; why can’t I die?” that same model would conclude the lifespan is too long.
I didn’t say that lifespan is too low, I said that the lifespan that you can choose if you wish is too low. The existence of people who want to die is irrelevant to this.
If we had lots of people lying around saying “I wish I could die; why can’t I die?” that same model would conclude the lifespan is too long.
We actually do have people around who want to die. At the same time we still want to increase the lifespan that people can achieve if they want to do so.
SilasX
Hmmm… a Bayesian optimization model will detect high values for a target function while remaining ignorant of very low ones. So I shouldn’t trust it?
Your optimizer, whether Bayesian or not, needs to be able to recognize a low point when it hits one, or else it can’t optimize at all! If every point looks the same… (It may learn more about high points, but it must still learn about low points.)
That’s not how Bayesian optimization works. Broadly, the idea is that we use Bayesian optimization when both calculating the value of the target function at a point and calculating its gradient are both expensive or infeasible. Thus, we instead choose points at which to sample the target function, and the samples train a Gaussian process model (or other nonparametric model of functions) that tells us what the function’s surface looks like. In such a procedure, we obtain the best performance by sampling points where either the expected function value or the model’s variance is particularly high. Thus, we choose points that we know are good, or points where we’re very uncertain, but we never specifically search for low points. We’ll probably encounter some when sampling points of great uncertainty, but we didn’t specifically seek them out.
I think that the lifespan that humans can live to if they wish, given current medical and scientific knowledge, is too low.
I agree.
The model I use to derive that involves looking at lots of dying people who don’t want to die. If we had lots of people lying around saying “I wish I could die; why can’t I die?” that same model would conclude the lifespan is too long.
I didn’t say that lifespan is too low, I said that the lifespan that you can choose if you wish is too low. The existence of people who want to die is irrelevant to this.
I think Sister Y would disagree with the implication that there are few such people.
We actually do have people around who want to die. At the same time we still want to increase the lifespan that people can achieve if they want to do so.
I think the key word in the part of dspeyer’s comment that you quoted is “lots”.