I don't know what is up with 1-sided score vs 1-sided approval. That is an interesting observation you made and does look suspicious; we should expect the distortion of power to be greater than the loss of honest expressiveness.
When I looked at this project a year ago, I found other strange behaviors that led me to ignore its results, similar to how you are comparing its graphs to your intuition. There is no "correct" model in voting theory. Models are neither good or bad, only to the author's taste or not. For example, with voters on a single dimension, should utility of a candidate be the squared distance, or the absolute distance? How should a candidate's positions be distributed, is n-dimensional Euclidean space good? Should some candidates be straight-up better than others, and what should the distribution be? How should voters be clustered, if at all? What does "strategic" voting mean in each context? What does "honest" voting mean in cardinal contexts?
The author's discussion of methodology was lacking, and the missing priors make its results partially meaningless. I also had no success looking through the code for even basic behavior, because it was too hard to navigate. However, this is only a graduate project, and most graduate projects (including mine) are of lower quality than this one, so I wasn't upset.
In score systems, voters can sometimes be made honest to an extent. For example, on a book rating site, you probably don't rate all books 0 or 5; you use the full scale. Your ratings have greater self-value than your desire to tip the global score. While your assertion is reasonable that it is not the expected "honest" voting behavior, it is also a behavior worth looking at. This goes back to discussion of methodology; all these assumptions need to be documented, rather than buried in hard-to-read code.
Intuitive expectations are not the same as empirical data.
> There is no "correct" model in voting theory.
Of course there is. The ideal would be a model that is exactly like real life.
> What does "honest" voting mean in cardinal contexts?
Normalization of sincere utilities.
> and the missing priors make its results partially meaningless.
Be specific. It just sounds like you don't understand how it was done.
Sure, and I can't criticize the model with any certainty, or make my own firm claims, without doing modelling myself. But looking for concerns is fine. I can't say the model is wrong in this aspect even though it looks odd. However, I can think of alternative models which would definitely have the opposite result. (For example, squared distance as utility.)
>Of course there is. The ideal would be a model that is exactly like real life.
Practical considerations make this impossible. Validation of a model by comparing to reality only works for small subsets of voting theory at the moment, and weakly so. A complete model like this one has no hope. And if comparison to reality is the goal, I can already see gaps where it diverges from the best-known theory, by omission of features. So taste is important. I have my own preferences about what features are important, which others will disagree with. These preferences will be non-falsifiable for at least 50 years while we wait for experiments to catch up.
For example, I assume you mean affine scaling of utility into the [0, 1] range when you say "Normalization of sincere utilities". But we might add to each election a candidate whose goal is to nuke the country itself. His utility is massively negative for every voter. However, it isn't to my taste for this candidate's presence to distort "honest" votes; I'd rather there be some clamping at the ends. So even this "obvious" choice is not canonical.
>Be specific. It just sounds like you don't understand how it was done.
Yes, that's the problem. I shouldn't have to go through code to understand necessary things. An expert in the field should be able to follow along by reading the paper. Although, this is not a paper, so holding it up to that standard is unfair. But the gaps in what I can figure out make the results non-interpretable in some of the aspects I most care about.
I think one of the results you most care about, from the model, is that Approval measures better than ranked choice. I think that's still a fair conclusion to arrive at.
> VSE cannot measure:
> - Any desirable characteristics of a voting method
> that do not directly relate to outcome (who wins).
Yes, this is how the terms are generally used in voting systems literature.
If I understand, "honest" voters in the score-like systems stick to some magically-shared mapping of their own utility to scores on the ballot (totally unreasonable, but let's grant it). Strategic voters, meanwhile, just use the ballots to maximize their own utility.
Let's take the simplest, common scenario of two clear frontrunners. In score voting's one-sided strategic scenarios: - Honest voters who support candidate A will put down magical normally-distributed numbers. Their scores for A end up perhaps 10-50% higher than their scores for B. - Strategic voters who support candidate B will maximize B's chance of winning by putting down the highest score for B and the lowest score for A. Score for B will be uniformly 100% higher than scores for B.
In this scenario it takes perhaps 2-10 honest A voters for every strategic B voter for A to win. Yet somehow, the VSE for 1-sided strategic score scenarios are all much better than plurality, and equal with approval?
This is a commonly-discussed downside for any score voting system, where it gives strategic voters extra power, and, in these 1-sided scenarios, is worse than plurality. The downside is appearing nowhere in this data.