Slashdot Mirror

Plurality is consistently bad. AV is somewhat less bad until it blows up spectacularly. Which is better is a matter of debate, but I'm not surprised that the voters would adopt a "better the devil we know" position. It's also rather telling that the Cons-Lib coalition neither went straight to MMP or STV nor ran a New Zealand-style two-stage referendum.

IRV sounds good, but it's actually rather bad in some non-obvious ways.

(I apologize in advance for the wall o' text, but I don't like linking to things without explaining them. There are explanations on the linked pages, so if you want to just click the links in order, you'll figure it out. I just like my explanations better...)

Yee diagrams are the simplest way to see this. They work by modeling a population of voters and candidates as occupying a two-dimensional issue space, say from -1 to +1 in each direction. (Of course higher and lower dimensions are possible, but 2D makes nice graphics that can easily be comprehended, and is high enough to demonstrate the pathologies and good behavior that apply in general.) Let N candidates occupy discrete points in the issue space, and are identified by unique colors; the voters are assumed to be distributed in a gaussian distribution about some point X in the issue space. (Yes, that's an unrealistic assumption -- voters are likely to be clumped in more interesting distributions. But if a voting system behaves for a simple gaussian distribution of voters about some middle position, it might work well for more complex cases -- if it breaks down into a pathological mess in the simple case, you know it's worthless for the complicated case.) Finally, assume each voter prefers candidates in order of their distance from his position -- and for voting methods involving strength of preferences in addition to ordering, preferences are a decreasing function of distance.

Now you run a simulated election for some large number of voters for each possible X -- that is, for every point in the rectangle from (-1,-1) to (+1,+1), at some discrete interval of your choosing. Each election is simulated, and the pixel at X is colored with the winning candidate's color. The candidate closest to X should win. (Obvious, right?) So the resulting diagram, for a good election method, should be composed of one contiguous, convex region around each candidate, matching the color of that candidate. (For the most extreme candidate in any direction, the region will extend to the edge of the diagram). AKA a Voronoi diagram, if you know what that is.

You can select the candidates' positions carefully to make a point (e.g. a simple case of 3 or 4 equidistant candidates, where one can intuitively predict the exact regions of an ideal election), or randomly. Yee's site has some of both, and I think you'll be surprised how IRV compares.

More at rangevoting.org for several methods and particularly for IRV. Note that the ones at rangevoting.org use fewer voters than Yee's simulations, thus resulting in more close-call elections going "the wrong way" compared to the limit with infinite voters following the same Gaussian distribution. While this may seem like a bug, it's actually a feature, as you can still mostly see the effects for large populations, but the stippled regions where some elections going either way reveal the relative sensitivity of voting methods, and thus which ones behave worse in real-life elections with small numbers of voters (e.g. county or town elections). While misbehaving for 5000-voter elections doesn't disqualify a method for federal or state-wide elections, all else being equal we'd certainly like to be able to use the same system for local, state and federal elections.

More relevant to reality, we can model more complex, and more realistic, voter populations, but since they can no longer be represented by a single coordinate pair, we can't just run it out into a intuitively readable diagram. But we can tally up the aggregate cost to society of each election result -- each voter suffers proportionately to the distance between the elected candidate's position and his own (or according to some other function than a n-dimensional isssue space -- e.g. some voters might

Didn't Eric Drexler and a bunch of other people use Crit Link or something similar to do this on a website my memory says it was edge.org in the late 90s/early '00s?

Don't get me wrong, a good-enough plan that's moving forward beats a perfect plan stalled in its tracks, but...

Do you realize just how pathological IRV is? Are you sure it's really good-enough, that a few publicized failures due entirely to the mechanics of IRV won't poison the voting-reform well for decades?

I like full-on range voting for its expressiveness, though approval voting (the 1-bit version of range voting) is nearly as good practically, but I'm not married to them -- if you value, say, moderation* over range voting's slavish representation of actual preferences, then pick a method like Borda that's ill-behaved (particularly, Borda is weak with regard to cloning) but biased in favor of moderates, not IRV, which (in addition to its nonmotonic pathologies), is biased in favor of extremists (if slightly less so than plurality/FPTP). See these Monte Carlo simulations of several voting methods.

  *Fundamentally, any bias is a risk -- if your voting system prefers moderates, a political machine can set up straw candidates to pull an Overton window stretch, not on the voters, but on the election itself.

I disagree.

Check out this site: http://zesty.ca/voting/sim/

IRV is only a little better than Plurality voting (what we have). It still favors the extremist candidates, and shuts out the moderates usually. IRV gives moderates just a little better chance to win, but not that much, and with some really weird behavior too where public opinion siding too much with a candidate will cause him to NOT be elected, and opinion swaying away from him will cause him to be elected.

The Borda system is the best, for making sure that the most centrist candidate is elected, but the Approval system isn't far behind it.

Sure, the Republicans might be thinking this will get them more votes from the third parties, but what will really happen is that the most moderate candidates will be elected. Because, while the Libertarians and NL people vote also for the Republican candidate, the Republican voters will also be voting for one or both of those 3rd party candidates too, and it'll really come down to 1) who's the most moderate of the bunch, and 2) where is the people's will--if it's closer to one of the moderates, then he'll be elected. These days, the Republicans seem to be pretty extremist.

Additionally, it could bring out candidates who, in the past, wouldn't have bothered to run, because they're too moderate and know the Plurality system prevents them from winning.

Alternative Vote can behave extremely oddly. Moving a candidate towards the top on your ballot can cause him to lose, and the Alternative Vote can also neglect to pick a candidate where a (different) majority prefers that candidate to each other candidate. The latter is what happened in the Burlington, Vermont election of 2009, and might have led the different majorities to unite against IRV (the Alternative Vote).

You can't say what would happen in a plurality election based on votes cast in an IRV election. In a plurality with 3 or more candidates, people vote tactically. You can't assume that their single vote would be the same as their first choice in an IRV election.

Alright, but that has no bearing as to the success of AV. If we disregard Plurality, the fact remains that Montroll is both the one-on-one winner and the one with the most first + second place ballots combined. My point was to show that AV is pretty unique in electing Kiss here. Of course you could say that that's not an argument, either, but in the worst case, we can disregard Plurality as it doesn't factor into the IRV/AV vs other methods comparison.

And what does "the worst candidate" mean anyway?

He loses to the other two. If it had been a top two runoff, he would lose no matter who the other candidate had been.

Again, that's crooked thinking. Imagine A is a left wing candidate, but the right has a split vote with candidates B and C both vying for the right. Say that the district is a right wing one. IT may well be that candidate A can individually beat both B and C. But the ideal vote is clearly not to return the left wing candidate for the right wing district. So again you can't just take the IRV votes and munge them into theoretical 2 horse races. It's garbage in, garbage out.

For your first point, consider a left-right spectrum with voters placed along the line and candidates at points close to some of the voters. Then start by "canceling out" the rightmost and the leftmost voter. Do so until there's only one voter left, and the candidate closest to him wins - in other words, the candidate closest to the median voter wins*. If a method satisfies the Condorcet criterion (i.e. picks the candidate that wins all pairups when there is one), then if voters vote honestly (vote candidates closer to them above candidates further out), the candidate closest to the median voter wins. That is called the Black single-peakedness theorem, and implies that a right-wing candidate would win in your example.
If you think about it, it makes sense: if the single left-wing candidate A is to beat both B and C, that means either that a majority puts A ahead of both B and C, in which case it's not a right-wing state, or that some voters rank "B above A above C" or "C above A above B", in which case right versus left wing isn't all that matters.
In a more general multidimensional case, the independence of clones criterion is a good yardstick for whether the method is vulnerable to vote-splitting or not. AV passes it, but so do the advanced Condorcet methods like Schulze. Yee diagrams also show that Condorcet methods find the median voter candidate in 2D instances if the voters are distributed in a Gaussian manner around some central point and the distance metric is Euclidean, although that's a more specific case.

For your second point, note that the pairwise margins are significant. If the voters were aiming to bias AV maximally, the margins would be tighter since the final comparison, when there are only two candidates left, is a pairwise one. Moreover, the fact that the AV opposition was able to unify when AV uniquely elected Kiss, but not in the prior election when it picked the same result as the other methods, suggest that AV did pick the wrong winner - that if there was strategy, it backfired. That is no proof, of course, but it does support the margins conclusion. Finally, we may add the first+second vote metric to the heap; if there was strategy, it would be unlikely that the strategy would show a clear winner by both measures, because if each voter crafts his vote to maximally affect the AV outcome, there would be few degrees of freedom left to make someone else the Condorcet winner and Bucklin (first plus second) winner as well.

* This is a reasonable g

Thanks for sharing. IRV is actually a pretty good system... in that it's better than plurality. In comparison, it could be better. Maybe there are manual count methods for better systems?

I have to say I'm still not entirely sure on the matter of electronic voting v. manual voting. I think most people who have strong opinions are basing them on too few criteria.

Well, somebody's gotta run in 2012, and most of the likely candidates are not too appealing. Not sure about that name though: "President eth1" doesn't exactly roll off the tongue.

Some thoughts on your agenda:

- Constitutional amendment: single-issue bills only.

I support the spirit of this proposal, but I worry that such an amendment would necessarily be so vague as to be easily abused. Who defines what an "issue" is? It's easy to imagine Congress defining "issue" very broadly and continuing to pass their over-9000-page porxtravaganzas, and one can also imagine a court defining issue very "narrowly" and striking down otherwise reasonable laws.

- Constitutional amendment: 10 year sunset clause on ALL federal laws.

The automatic-sunset idea is intriguing, but it's also prone to abuse. We would probably just acquire a new tradition, wherein a whole slew of laws are rubber-stamped for renewal on the first day of each Congress, with the only results being that some junior members get gavel practice and the poor President gets writer's cramp.

(create an upper bound on the number of laws that the federal gov't can maintain)

How do you choose what the upper bound should be? And what happens when the Elbonians invade and Congress can't declare war because they're already at quota? I tend to favor the approach of just sticking to the enumerated powers, although admittedly that hasn't worked out as well as one might have hoped.

- Move elections to an instant-run-off system so voters don't feel they have to try to game the system

I believe the advocates of instant runoff voting have the best of intentions but are betting on the wrong horse. IRV is the only widely proposed voting system that is arguably worse than our current system, and certainly it won't eliminate gaming the system. In my book, range voting is the best system, and approval voting is nearly as good, with the added bonus that it's very easy to understand and wouldn't require changing ballot designs (which could be relevant to persuading people to accept a change). For those who may be interested, Wikipedia has a pretty good set of articles, and check out these pretty pictures of the bizarre things that can happen under IRV.

- Move election day to July 4th. More people vote because they're off work. Can celebrate *getting* freedom and *keeping* it.

Of course they're also on vacation, at barbecues, eating dozens of hot dogs, shooting off fireworks, etc. Many will be too busy loving America by means of combustible projectiles to love America by means of throwing the bums out.

At Pycon 2007 in Dallas I saw a lightning talk demonstrating how electronic voting could be secure in just 293 lines of code.

However the bottom line is that you shouldn't trust any voting system. What you should have is an auditing system where you can do recounts. The less moving parts or the fewer lines of code you have the easier it will be to audit a system.

Granted, each voting method has its own flaws -- Kenneth Arrow showed that not every desirable voting criterion could be satisfied simultaneously -- but that does not mean all methods are equally bad.

Rather than use a lot of boring math, it's easier to show with pretty pictures: http://zesty.ca/voting/sim/

This shows voting simulations using Plurality, Borda, Approval, Condorcet, and IRV. (Note, the reason I left Borda out of the post above is because it seems to be more sensitive to strategically burying opposing candidates).

You don't identify yourself to a car, you simply use a key (password), the car doesn't care who you are, nor does it need to. You can give out copies of the keys or loan the keys to someone else so they can drive your car. Same goes for the combination to a combination padlock for a bicycle-- the lock doesn't care who anyone is. Have I understood what you're getting at, naasking? Capability based security is much like password only access? Like keys to a car?

You pretty much have it. The "capabilities as keys" analogy has been used extensively before, and its accurate to a first approximation. Capabilities have stronger security properties than keys though, so don't take the metaphor too far. Here's a paper discussing the various forms of capabilities and debunking the myths promulgated about them over the years.

And using only a password to authenticate is also a good idea, provided the password is a cryptographically secure identifier.

No, IRV does not eliminate spoilers. The spoiler effect exists in IRV. With IRV there are still wasted votes, except it's much harder to tell which votes will be wasted and what effect your ballot will have; the behaviour of IRV is much more complicated, often pathological, and thus arguably even worse than the current system.

Range or approval voting would be a better option. They truly eliminate the spoiler effect, they are easy to implement using unaltered existing equipment, and have simple, easily understandable behaviour.

Instant runoff voting is non-monotonic, which means you can cause your favorite candidate to lose by voting for him/her. There are a host of other problems with it. Promote approval/range voting or a Condorcet method instead.

A great paper summarizing these issues:

http://zesty.ca/lj/yee-oca-transferable-vote-3.pdf

http://zesty.ca/voting/sim/ illustrates clearly some of the really wacky things that can happen with IRV. For example, in certain situations, a shift in the public opinion towards a certain candidate can actually cause that candidate to lose.

Here is an interesting site which gives you some visualizations of voting methods. Take a look at IRV:

http://zesty.ca/voting/sim/

The vehicular anaalogy is good because it illustrates the flaw in the current security systems.
The current system amounts to driving blind.
When I load an email virus it runs with all my authority or not at all.
That is like being placed into my tank with a blindfold on and being that I can go forward or not and whatever happens is my fault.
This is crazy I am not going to drive with a blindfold on.
TAKE THE BLINDFOLD OFF.

see http://zesty.ca/capmyths/

But how to get there?
All of our current systems make use of static security and not dynamic, eyes open, security.
I think the virtualization will offer a way out.
One machine running a secure, capabilities based OS and other machines for legacy applications.
Can anyone say L4 with capabilities? ;-)

Security is all about preventing undesirable events.

Even if we add a wonderful UI to security, it will never be perfect.

Security is about saying "no", ease of use is about saying "yes."

Instant runoff voting does not solve this problem. The IRV system is widely misunderstood.

1. IRV only allows you to safely vote for a third party as long as the third party has minimal support. Once the third party begins to gather substantial support, IRV prevents it from winning exactly as the current system does.

2. In many realistic situations, IRV exhibits crazy behaviour in which voting for a candidate can cause him to lose. No other seriously proposed voting system has this problem.

3. IRV is complicated to explain and expensive to implement. The ballots cannot be totalled up; instead, information about every individual ballot must be sent to a central location to determine the winner.

Approval voting has none of these problems. In approval voting, you vote for as many candidates as you want, and the candidate with the most votes wins. It is extremely simple to implement, easy to understand, works with existing ballots, and truly allows you to vote for any party you like because all your choices are independent.

See this document for a detailed comparison.

I agree with you that capabilities are great, EROS has the right approach, and NGSCB is the wrong approach. Alas, SELinux does not have "capabilities" in the EROS sense.

Unfortunately, the term "capabilities" was misapplied to "POSIX capabilities", which are what SELinux has. These are not capabilities at all in the original sense of the word (the sense they are used in EROS). For a detailed explanation of the differences, please see http://zesty.ca/capmyths/usenix.pdf.

I'm not sure if MS aren't talking about something different from what most of this discussion thinks they are. Rather than showing the thread of discussion of whole emails (which we're all used to in other clients) it might be they mean something more like this old discussion of what e-mail discussions should look like by Ka Ping-Yee..

In case you manage to /. that, the idea is that it shows the responses to pieces of your email - the kind where someone says "see my responses inline" and responds to each of your points piecemeal, then you do the same to their responses, and so on.

I've often thought it would be cool to write something to parse emails the KPY way, but the heuristics would have to be pretty damn clever to deal with supercite. Specifically what I wanted was something that combined KPY's ideas with text-autosummarization , and some 'author ranking' information to produce mailing list summaries from gmane which are like Kernel Traffic and Cousins, or the now-defunct Eclectic.

Oh well, I can always wait until MS put this in Outlook 2010 ;)

That being said; the Roundup I tested and tried was (and still appears to be) very rough. Is yours a continuation of that effort? Or a fork? Or a similar project with the same name? There isn't mention of the relationship on Ka-Ping's page.

Regarding the Python statement, it's simple; Python is a userspace binary which is called at each load of each page (evidenced by watching the processes in real-time on the server as hits are being made), whereas PHP lives inside the namespace of the webserver process itself. I'm also good friends with the author of PHP, so I can get direct help at a moment's notice when I run into snags.

I realize that I can probably load mod_python into the server as well (DSO), but at the time I was testing, mod_python wasn't mature or functional enough to be usable in a production environment. It really boiled down to that.

And at this point, we have so much metadata and bugs in the system across several dozen projects, I'm not sure switching would be easy, unless Roundup (your version) can cleanly import all of the metadata that Mantis manages, from users to passwords to bug history and file attachments. A long shot, to be sure, but I'll reconsider Roundup again in the next round of evaluations for new projects. It might make a good option for users to pick in the ala-carte menu at project signup time.

Thanks again for the comments and feedback; always welcome.

When i started working as a developer at ILM, there wasn't really any bug-tracking system in use. So i threw together a quick hack, on a weekend when i finally couldn't stand it any more (isn't that how so many software projects get started?).

Over the two years i was there, it grew, but it stuck to three core ideals:

1. Don't force the person entering the bug to enter anything other than a description. Just allow people to go back later and add details, change categories or priority settings if they want to.
2. Make it really easy for other developers to enter the conversation around the bug, so the discussion and activity surrounding the bug can be recorded with the bug.
3. Optimize the user interface to display the most useful information in the available screen space.

To meet the first two goals, the system was based on e-mail: anyone could just send a free-form e-mail message to the roundup address, and a new bug would be created; then anyone who replied to messages about the bug would be automatically added to a mini-mailing-list. Every bug got its own automatically managed mini-mailing-list.

Most other bug trackers are really bad at the third goal: they take up all kinds of space with management details that people don't really care about. What a developer really wants to see is the descriptions of the bugs, so Roundup maximizes the screen space for that.

It grew far beyond my original plans, in terms of the number of users and bugs logged. The back-end implementation was terrible (no database, just lots of little text files; performance was awful but it did have the redeeming factor that you could just use "tar" to archive them).

As far as i know, ILM R&D is still using it today, and they're running multiple instances of it to support different applications and teams.

You can get the source code and see some screenshots on a page about Roundup, though i'm not developing it any more. A new group of people has picked up the torch and carried it on in a Sourceforge project that is alive and well. Their project is a complete rewrite of Roundup, originally based on a design document I wrote, but now much extended. I encourage you to check it out.

There's a short paper explaining this system at http://zesty.ca/crit/yee-crit-cscw2002-demo.pdf.

There's a short paper explaining this system at http://zesty.ca/crit/yee-crit-cscw2002-demo.pdf.