STAR vs. Score
STAR vs. Score
Is there an example, starting with hypothetical voters' true valuations of the candidates, and using strategy, where these systems don't produce the same outcome? If not, there are no grounds on which to justify the extra complexity of STAR.
@Jack-Waugh Jameson Quinn did a simulation which included strategy and STAR came out ahead. I would tend to break it down that score is better with 100% rational, 100% honest and 100% informed voters and STAR is better otherwise. Score also can suffer from an issue if there are no good candidates of a large portion of voters.
For myself, I have one issue with STAR. It is majoritarian. For this reason I invented a new system
It does not have that issue but is more complicated. I doubt that it is a net positive trade off since STAR is simple enough to explain to a layperson very quickly and STLR is not. This means that it is not viable in a referendum.
@Keith Do you know what assumption Jameson Quinn incorporated as to strategy with Score?
@Keith Maybe you are doing it right, and the software isn't. Or maybe the designers of it decided not to indent.
@Keith The word "strategy" does not appear in https://www.equal.vote/science . If I can't see an example that shows the sentiments of the hypothetical voters, and what strategies they used, I do not have before me an example that I can study.
To me it is paradoxical that two systems that meet the balance condition can produce different results from the same electorate (once they understand what strategy is most effective).
@Jameson-Quinn, in your VSE sim, what did you assume as to what strategy voters would use with Score?
@Jack-Waugh The word strategy appears several times on that page. I do not now exactly how he coded the different strategies but I figure that could help you on your way to doing research.
I am unaware of any property called "the balance condition" and electowiki does not have such a page. Do you intend to refer to The Test of Balance given here. If so I think what you are saying is that in Nash equilibrium two systems which pass this criteria should behave such that the strategy of different factions cancel each other out and they both produce the same winner. The flaw in that logic is the assumption of an underlying symmetry in the size of groups and how that interacts with compromise/utilitarian or majoritarian winners. As I said above, STAR is majoritarian and Score is Utilitarian. In the absence of strategy these systems will give different winners. So even if all the strategy cancelled you would not expect the same winners.
@Keith, by "the balance condition", I meant Frohnmayer balance.
@Keith, I was not thinking anything about a Nash equilibrium. I do not know what that is.
@Jack-Waugh Yes that is the same concept as I gave the link to. My comments still apply.
Voting without strategies, range [0,9]:
A B C
The voter thinks that B and C are the 2 frontrunners, therefore:
SV: A B C
Whether with the min-max strategy, or without, the voter wants his vote to be worth the maximum in the clash between B and C (certainly not B C).
STAR: A B C
In the clash between only B and C, the vote would automatically become B C, so the voter does not need to lie at the beginning.
If he uses min-max the vote becomes:
A B C which is still better than SV.
Votes without strategy like this:
55%: A B C
45%: A B C
STAR wins A while SV wins B, with almost double the points of A.
In a strategic min-max context (and forecast on frontrunners) the votes would become as follows:
55%: A B C
45%: A B C
and A would also win in SV.
The point is that we should make B win as much as possible in a similar context and with STAR, B certainly never wins (strategies or not).
The practical example would actually be:
55%: A > B
45%: A < B
with A and B finalists. The individual ratings of A and B can change but in the majority methods, A always wins, while in the utilitarian ones, B can win when it has greater utility.
Other methods such as STLR and DV avoid the SV defect, without however being majoritarian like STAR. However, they can have other flaws (es. STLR can lose its positive sides in the presence of clones making STAR a better alternative for semplicity, DV fails monotony).
STAR does still suffer from some of the same drawbacks as Score, mainly that it is majoritarian, although it does manage to approach a consensus model more closely and to reduce the effectiveness of strategy. I do think that STAR is a pretty good system and that it is superior to Score.
One modification I would make to STAR is something I have advocated before, which is an exponential demerit score system. Basically, there are N score choices to give to each candidate, each score option being an order of magnitude higher than the next smallest. Candidates are approved by a lack of demerits, and the two candidates with the fewest demerits have a runoff between them based on which was scored more favorably than the other more often. The runoff is for the same reason as in STAR while the exponential demerit system empowers a broader consensus over a majoritarian front.
I wonder if there is a systematic way to vary the weight between consensus candidates and majority candidates. For example, a candidate can be called (S,P)-consensual if at least a P-fraction of voters scored that candidate at least an S. If you plot all (S,P) such that there is an (S,P)-consensual candidate, you can choose a method to try to trade-off S versus P in a controlled way. For example, you could select a candidate with a minimum value of
Really any reasonable metric could do.
@cfrank Score is not Majoritarian. It is Utilitarian.
Score is Utilitarian but does not adjust voter impact to reduce strategic incentives. STAR is Majoritarian but does adjust voter impact to reduce strategic incentives. I played with this issue a lot and came to the compromise that STLR voting was the best tradeoff.
Its also worth noting that some people actually prefer Majoritarian systems and think that the tyranny of the majority is justified. Many of these people are the IRV supporters. This makes STAR more desirable to them and therefor a good system to campaign for strategically.
I think all the replies to date on this topic are from people who either, like me, have skepticism about STAR, or who outright oppose it on various grounds. The STAR advocates haven't yet spoken up. I want one of them to present an example where the results differ and say that they think STAR produced better voter satisfaction overall than Score would have. I expect I will be able to counter that when they worked their example, they did not use the optimal strategy with Score. So the end product I expect to happen from such exchanges is a lack of cases that count for STAR. If the outcomes aren't any better, there is no justification for the extra complexity.
STAR is proposed as an "IRV 2.0." In response I propose an "IRV 3.0," which has more to do with IRV, because it accepts votes in IRV style, which STAR does not.
@Keith Yes that is true, although I don't really understand why people would be content with tyranny by the majority when it may be possible to avoid it (unless they are a part of the majority). I'm actually not much of a fan of utilitarianism, I am in favor of something that is distributionally just. STLR voting is interesting.
@Jack-Waugh, in my opinion STAR is way better than IRV. I'm not exactly an advocate of STAR, but I think in general that rank-order voting is probably not going to be a great solution. I think that "independent" scores is really the best I think we can hope for for now, and that means it's what we do with those scores and how that interacts with voters' decision-making that's important.
@Jack-Waugh What STAR does is it renormalizes everybodies vote weight to give them the same impact. This is an attempt to reduce the amount of strategy needed. I do not think that it would outperform somebody who used optimal strategy with score. The point is that most people do not or cannot use optimal strategy. STAR then puts people bad at strategy on a closer level to those who are good at strategy. So I do not think you are wrong in what you say. If all people where fully informed, rational and strategic then score would likely be better. However, people are not any of those things in general. I do not think your like of argument will hold up under this consideration.
An example of where score produces a better outcome than score is
40% = A:5 B:0 C:0
31% = A:0 B:5 C:1
29% = A:0 B:1 C:5
Score give A and STAR gives B. This is an engineered and somewhat extreme example to illustrate the issue. Is 5 infinitely more than 0 or just 5. Is 5 weighted as 4 more than 1 or 5 times. There is no universal metric and different people will choose different metrics. STAR normalizes it all away and compares the two most favoured with full weight to each voter.
STAR is a simplified version of Baldwin's Method. When you think about it that way you see the intent.
Given these 3 types of ratings (assuming they are the ratings of the 2 frontrunners, after eliminating all the others):
[0,1] - [2,3] - [4,5]
STLR normalizes them like this:
[0,5] - [3.33,5] - [4,5]
Baldwin normalizes them like this:
[0,5] - [0,5] - [0,5]
For me, STLR uses better normalization but I don't think it's the best.
If a vote like this: [4,5] remain the same in the clash between the two finalists, the voter from the start will be encouraged to downplay the rating of the worst candidate of the 2 (i.e., to vote like this from the start [0,5] ).
I prefer this normalization in clash between two finalists:
- if you have a couple [0,0] or [5,5] the vote is irrelevant.
- if one of the two candidates has a score of 5, the other is put at 0.
- if one of the two candidates has a score of 0, the other is put at 5.
- if both candidates have intermediate scores, then STLR normalization applies.
For simplicity, I call START the STAR that uses this normalization.
In this way, at the beginning the voter:
- first assigns 5 to his most favorite candidates and 0 to the most hated ones.
- then he can feel freer in assigning intermediate scores.
Such normalization is proposed indirectly in Tragni's method, although in that context it is used to make comparisons between couples.