So my slightly crazy article about random ballot was mentioned in the new scientist.
This is, as you might expect, quite exciting. Unfortunately the article gets one really crucial thing about me wrong, and my comment clarifying this on the site has yet to be approved, so I just want to make this clear:
I am not actually advocating random ballot
Observe the following quote from the original article:
I think it has desirable and undesirable features (in particular there’s one really major problem with it), but even with the above list I’m not sure I actually endorse it. Consider this a thought experiment more than a proposal.
“perfect” was in scare quotes for a reason. It has some very nice properties, but it also has a lot of problems (both theoretical and practical). I have no interest in promoting it as the system we should actually be using, it’s just an interesting system to think about it. The reality is that (while it’s not necessarily the system I would have most preferred) I am extremely strongly in favour of the move to AV and will be voting yes on May 5th.
Update: Jacob has now fixed this, mentioning I am “highlighting” it rather than “pushing” it.
At the risk of stating the obvious, the random ballot idea is basically the same as a colour dithering algorithm.
In fact if you imagine a map of the UK with each constituency as a pixel in a RGB colour giving the proportion of votes (R = Labour, G = LibDem, B = Conservative), then it’s like an algorithm that’s only allowed to use pure R, G or B but tries to give the right shade.
Hmm, what about starting with an actual dithering algorithm – say one that compensates for an inaccurate colour in one pixel by adjusting neighbouring pixels – and converting it to a voting system? Can’t believe this hasn’t been thought of before.
I like that. I wonder how well it could be made to work.
The desire (in some voting systems) to make the proportion of MPs in a region roughly correct, like multi-member constituencies, seems very similar to the desire for colour dithering to make a shade correct locally even though no 1 pixel can be exactly right.
Yeah, I understood the reasoning. I was just wondering how it would look in practice. It’s made more complicated by the fact that the “pixels” are not square and butt up against eachother in odd ways.
I know almost nothing about colour dithering algorithms. Could you sketch how it might be applied here?
> I know almost nothing about colour dithering algorithms. Could you sketch how it might be applied here?
Nope cos I know very little about them either (not enough to speak authoratitively on the subject), but I don’t doubt that the principle could be made to apply to voting. If only because the goals are basically the same – the local political ‘colour’ should be as close as possible on average even though you’re forced to quantise it.
The perils of online notoriety David… :-)
Very enjoyable series of posts!
David,
Actually I missed that quote. That changes my impression of your contribution to the article considerably. I apologize for unfairly judging you.
I don’t think that should cast a negative impression of Score/Approval Voting advocates in general however. This was just my own personal error. Nor does my mistake trump the massive evidence in favor of Score/Approval Voting, for the specific case of single-winner elections.
For multi-winner, Reweighted Range Voting and Asset Voting are nice.
I should note that I very much don’t advocate random ballot for single-winner elections. I still maintain the idea has merit for electing a sufficiently large house, but it would be sheerest madness to use it to e.g. elect a president. (Edit: I suppose technically electing a candidate is a single-winner election, but the point is that you only consider the elected candidates en masse)
Here are my objections to range and approval voting:
* Range voting I am undecided upon. I think some of the arguments for it are compelling, I think some of them are distinctly uncompelling. Additionally, experience of non-political voting systems is that most people when you ask them to provide a range will just degenerate into the two extremes – people find it hard to make up a score when there’s no actual meaning attached to the number. I am fully prepared to be proven wrong on this, but I consequently tend to feel that range voting is likely to degenerate to approval voting in practice. Which I disapprove of because…
* Approval voting is a *strict subset* of comparison based voting with a RON entry. If you treat RON as an implicit “everything I haven’t voted for is undesirable” then approval voting is a strict subset of partial ranked voting. You’re essentially throwing away a large amount of information about my preference as a voter, which is information you absolutely need in order to understand my actual opinion.
By the way, the “Range voting advocates are disconcerting people” tweet was more of a general impression and not actually about you.
> approval voting is a strict subset of partial ranked voting. You’re essentially throwing away a large amount of information about my preference as a voter, which is information you absolutely need in order to understand my actual opinion.
You’re making a common novice fallacy, explained in detail here:
http://www.electology.org/expressiveness
Long story short, the amount of information your ballot can hold is merely one consideration. It also matters how much of that information the system *uses* in the tabulation process, as well as how much inaccuracy the system contains due to tactical behavior.
E.g. there are 3,628,800 ways to rank 10 candidates, but only 1022 meaningful ways to approve them. That makes it seem like a ranked ballot is a lot more expressive than an Approval Voting ballot.
But consider that Instant Runoff Voting, for instance, ignores a massive fraction of the data that you put on your ballot:
http://scorevoting.net/IrvIgnoreExample.html
Also, virtually every ranked system fails the Favorite Betrayal Criterion, which means that your general best strategy is to switch your sincere ranking of ABCD to BADC if B and C are the presumed frontrunners. This is a massive distortion of true preferences, that also tends to lead to two-party domination:
http://ScoreVoting.net/NESD.html
Whereas tactical behavior with Approval Voting tends to elect Condorcet winners.
http://scorevoting.net/AppCW.html
This all comes through in Bayesian Regret calculations, which show you how representative each election system actually is, when you combine the information content of the ballot with the tactical and tabulation impacts I just mentioned. Score Voting and Approval Voting excel.
http://ScoreVoting.net/BayRegsFig.html
Moral: focus on Bayesian Regret. Focusing on individual properties is like debating race cars on the basis of their engines or aerodynamics rather than their average race times in timed trials.
> most people when you ask them to provide a range will just degenerate into the two extremes
Which is absolutely fine. As long as a significant *portion* use the intermediate options (and tons of experimental data says they will) you get better outcomes than with the already excellent Approval Voting.
> people find it hard to make up a score when there’s no actual meaning attached to the number
As someone whose conducted Score Voting exit polling as early as 2006, I can tell you that you’re dead wrong. For instance, I did this small poll in Beaumont, Texas:
http://scorevoting.net/Beaumont.html
Here’s a much larger experiment conducted in Orsay, France.
http://scorevoting.net/OrsayTable.html
Voters experimentally complete Score Voting ballots faster than ranked ballots. That makes perfect sense, since Score Voting is actually computationally simpler than ranking according to an objective metric called Kolmogorov complexity. The algorithm is simply to make one pass through the list to find the min and max utilities, and then make a second pass to normalize the utilities to that scale. For ranked systems, a full sort is required. Voters intuitively use a selection sort, which has order of complexity n^2—very inefficient for larger lists.