Interesting comparison of the effectiveness of different voting methods


In the study, different machine learning algorithms are trained to correctly classify an image, by assigning that image a score equal to the probability that that algorithm thinks the image is in a particular category for each category. The study is about when multiple machine learning algorithms give an image different scores, what way of combining the scores best increases the chance that an image will be correctly categorized. (I think).

In this study, I believe STV is just referring to single winner STV (aka IRV).

The sum rule is honest utility score voting where voters know what their own honest utilities are relative to everyone else, give their honest utilities on an infinite sliding scale style score ballot, and do not normalize their ballots.

The product rule is the basically the same, except for that honest utilities are log’ed before being summed together. Because the partial sums of the harmonic series approximate the natural log (since ln(x) = the integral of 1/x, and the partial sums of the harmonic series are just partial sums of the 1/n series), the product rule is most similar to harmonic voting (which **non-**sequential proportional approval voting is a special case of for when approval ballots are used) and psi voting.



Who approved those unreadable graphs? :roll_eyes: