Re:Why Hacker News should use a different Deterministic Algo...
by yummyfajitas on Monday December 02, 2013 @11:47AM
(#6834731)
I have little dispute with everything you've written.
However, one place I'd argue with you strongly is that handling beliefs about the world should be part of your model. For example, you mention different upvote probabilities on the new and front page - I'd simply model that with a new parameter beta (possibly related to alpha in some way).
It's far from clear to me how randomization such as what Dan Luu proposed actually addresses this at all.
I'll also discuss one other point: The author here suggests computing the posterior probability of an article being better than another for each pair of articles, which seems rather impractical.
It probably is. However, suppose we went through the effort of computing posteriors, only to discover that we get stuck at the stage of computing pairwise probabilities. There is still a very good chance we can draw samples from the posteriors.
Even though our deterministic algorithm has failed, we still get a much better random algorithm out of the exercise. Given the posteriors on P(good article), we can draw samples from them and employ Thompson Sampling to sort the articles. (I.e., draw a sample from each posterior and then order the articles by increasing sampled posterior.)
However, one place I'd argue with you strongly is that handling beliefs about the world should be part of your model. For example, you mention different upvote probabilities on the new and front page - I'd simply model that with a new parameter beta (possibly related to alpha in some way).
It's far from clear to me how randomization such as what Dan Luu proposed actually addresses this at all.
I'll also discuss one other point: The author here suggests computing the posterior probability of an article being better than another for each pair of articles, which seems rather impractical.
It probably is. However, suppose we went through the effort of computing posteriors, only to discover that we get stuck at the stage of computing pairwise probabilities. There is still a very good chance we can draw samples from the posteriors.
Even though our deterministic algorithm has failed, we still get a much better random algorithm out of the exercise. Given the posteriors on P(good article), we can draw samples from them and employ Thompson Sampling to sort the articles. (I.e., draw a sample from each posterior and then order the articles by increasing sampled posterior.)