So, “probabilistic constraint methods” is a trigram that I don’t think has an established unique technical meaning, and the various bigrams that can be formed from it don’t seem quite correct to me. I suppose the thing I’m objecting to the most is the absence of ‘sampling,’ but even then it has to be clear that the thing being sampled is points from the distribution, rather than constraints from the constraint set.
At least in the plate example, they aren’t “probabilistic constraints” in that the constraints are satisfied with some probability*--the way the method works is that they put a ‘uniform’ prior over ‘all’ layouts (that both satisfy the constraints and don’t), then update on the fact that the layout satisfied the constraints to get a posterior distribution. (The thing that seems to be specific to that paper is that they have a reasonable prior distribution over open layouts, which allows them to extend the method to that domain.)
(A related idea is also a thing in optimization—start with a prior distribution over feasible solutions, update with a likelihood function that’s the utility function to get a posterior, and then try to estimate the mode of the distribution using standard statistical tools like MCMC.)
eli_sennesh’s primary point, as far as I can tell, is that we should sample from the feasible region and then use our human judgment on a population of candidates, rather than trying to optimize using machine judgment and then only consider the one candidate it produces. But none of the three words of the trigram deal directly with that claim!
*Though the method can handle that gracefully, for the obvious reasons.
a trigram that I don’t think has an established unique technical meaning
For what it’s worth, that’s my impression too. So I take it Eli is coining his own term; I don’t see anything wrong with that.
I suppose the thing I’m objecting to the most is the absence of ‘sampling’
I take it you mean that you’d like the 3-word description to include “sampling”, rather than that the 3-word description implies sampling that isn’t being done (which is how I first misinterpreted your comment!). I agree that a description with the word “sampling” in might have been more informative—but probably necessarily longer too.
they aren’t “probabilistic constraints” in that the constraints are satisfied with some probability
I was parsing the phrase as (probabilistic (constraint methods)) rather than ((probabilistic constraint) methods) and therefore wasn’t expecting to see the constraints being satisfied only with some probability.
Anyway: It’s possible that Eli didn’t choose the best possible 3-word description for the class of methods he had in mind. But that seems a quite different complaint than that the paper doesn’t embody the term as Eli meant it.
So, “probabilistic constraint methods” is a trigram that I don’t think has an established unique technical meaning, and the various bigrams that can be formed from it don’t seem quite correct to me. I suppose the thing I’m objecting to the most is the absence of ‘sampling,’ but even then it has to be clear that the thing being sampled is points from the distribution, rather than constraints from the constraint set.
At least in the plate example, they aren’t “probabilistic constraints” in that the constraints are satisfied with some probability*--the way the method works is that they put a ‘uniform’ prior over ‘all’ layouts (that both satisfy the constraints and don’t), then update on the fact that the layout satisfied the constraints to get a posterior distribution. (The thing that seems to be specific to that paper is that they have a reasonable prior distribution over open layouts, which allows them to extend the method to that domain.)
(A related idea is also a thing in optimization—start with a prior distribution over feasible solutions, update with a likelihood function that’s the utility function to get a posterior, and then try to estimate the mode of the distribution using standard statistical tools like MCMC.)
eli_sennesh’s primary point, as far as I can tell, is that we should sample from the feasible region and then use our human judgment on a population of candidates, rather than trying to optimize using machine judgment and then only consider the one candidate it produces. But none of the three words of the trigram deal directly with that claim!
*Though the method can handle that gracefully, for the obvious reasons.
For what it’s worth, that’s my impression too. So I take it Eli is coining his own term; I don’t see anything wrong with that.
I take it you mean that you’d like the 3-word description to include “sampling”, rather than that the 3-word description implies sampling that isn’t being done (which is how I first misinterpreted your comment!). I agree that a description with the word “sampling” in might have been more informative—but probably necessarily longer too.
I was parsing the phrase as (probabilistic (constraint methods)) rather than ((probabilistic constraint) methods) and therefore wasn’t expecting to see the constraints being satisfied only with some probability.
Anyway: It’s possible that Eli didn’t choose the best possible 3-word description for the class of methods he had in mind. But that seems a quite different complaint than that the paper doesn’t embody the term as Eli meant it.