[EM] Realistic strategy questions
Kevin Venzke
stepjak at yahoo.fr
Mon Feb 24 12:30:42 PST 2014
Hi Toby,
In my own simulations (in which voting blocs learn how the method works through repeated polls, until ultimately one poll becomes the election result) the consistent difference between WV and margins was favorite betrayal (i.e. WV sees compression incentive where margins sees compromise incentive). IIRC the two usually scored pretty similarly with regard to sincere CW election rate and "utility maximizer" election rate.
Actually, I found most methods pretty hard to differentiate wrt social utility. Given the vast range of possible underlying voter utilities, I think the actual voter utilities are very hard to perceive, even if the (strategic) votes aren't very distorted from the truth.
Jameson, does your strategy #3 "fully strategic" mean Warren-style "use favorite betrayal and burial no matter what"? Or is it more of an approval-style "everybody gets an A or an F" thing, under the methods that allow it at least?
Does your polling system do a round of Range voting, or something like that?
My concern would of course be the possibility of supposing behavior that doesn't happen to make sense in a specific scenario. I see that you have some conditions to try to avoid this. But naturally if "strategic" voters strategize poorly (either too much or too little, or not in the right way) or "unevenly" poorly (among the various methods and scenarios) this could bias a method's final score (either positively or negatively!).
Kevin
________________________________
De : Toby Pereira <tdp201b at yahoo.co.uk>
À : electionscience at googlegroups.com
Cc : election-methods <election-methods at electorama.com>
Envoyé le : Lundi 24 février 2014 12h21
Objet : Re: [EM] Realistic strategy questions
Can I ask which voting systems will be included in this? Among other things I'd like to see how different Condorcet methods compare with each other and also margins v winning votes. Certainly something to look forward to, although I have nothing useful to add about the strategy types at the moment.
On Monday, 24 February 2014 17:46:39 UTC, Jameson Quinn wrote:
I have been working on my program for measuring the Voter Satisfaction Efficiency (formerly known as BR) of various systems. It's just about ready to do a big run, but before I do that, I'd like to get as many of the activists here to say what they think are reasonable settings for generating scenarios. In particular, I have to decide proportions for various kinds of strategic voters.
>
>
>Here's a few possible kinds of voters:
> 1. Honest: Will always map their utility onto a vote as honestly (linearly) as possible, after normalization over the range of candidates available.
>
> 2. Media-based Honest: As above, but when normalizing, they will ignore all candidates who are worse than both polled frontrunners.
>
> 3. Fully strategic: Will always strategize by maximizing the ballot distance between the two (honestly) polled frontrunners.
> 4. Weakly strategic: Will strategize, but only as much as "necessary". That is, assuming no other voters strategize, they will shift their vote, focusing on the margin between the polled winner and runner-up. In a Condorcet system, such a voter would always be honest. In a median system, they would exaggerate to ensure they graded the winner and runner up on opposite sides of both of their polled total grades, but not necessarily to an extreme; so if their honest vote was A,B,C,D,F and the polled grades were D,C+,C-,D,D, then they'd vote A,B,D,D,F.
>
> 5. Lazily strategic: Fully strategic if and only if their weakly strategic vote would not be identical to their honest vote.
>
> 6. Threshold strategic: Fully strategic if and only if their utility (satisfaction) difference between the polled frontrunners is greater than some threshold (in absolute value)
>
> 7. One-sided strategic: Will strategize if and only if they prefer the polled runner-up to the polled winner.
> 8. 20/20 hindsight: Fully strategic if and only if one of the last N election results was changed by strategy. (When simulating, you'd just use some constant probability for each election system, and find an equilibrium point for that probability).
>
> 9. Sheep: strategic iff more than X% of the non-sheep voters are.
> 10. Cliquish: Certain probability of being each of the above kinds, except with a certain extra probability of being the same as other voters who share similar utilities.
>Obviously, it would be easy to extend the above list by combining the various aspects there.
>
>
>Personally, I find ALL of the above strategic types to be essentially plausible. I know that the full decision rule for some (such as weakly and lazily) are more complex than the explicit thought processes of 99.99% of voters, but in practice I think it would be easy to end up acting like that through implicit and/or subconscious heuristics. Thus, in particular, I find it extremely IMplausible that the electorate would consist only of types 1 and 3, as previous BR simulations generally assumed.
>
>
>The question, then, is: how many of each kind of voter should I put in? Of course, I'll run several scenarios, but there's no way I can fully explore the 8-simplex of possible combinations of the above 9 voter types. So I have to choose where to focus my attention.
>
>
>And furthermore, I think it's valuable to say which percentages you find plausible before you learn how your favorite voting system does under those percentages. In my debugging so far, I've gotten a glimpse of the impact of types 1, 3, 4, 5, and 7 above; but I have not looked at the rest at all. So here's a scenario I find reasonable, and which I really don't know how it will turn out for my currently-favorite systems:
>
>
>Each voter has a "honesty type" which are 50% honest and 50% media-based; a "strategy type" which are 75% full and 25% weak; and a "decision type" which are 30% hindsight (N=1-3), 30% threshold (X=0.5-2 SD), 20% lazy, 10% one-sided, 5% always-honest, and 5% always-strategic. They use their "decision type" to decide whether to vote according to their "honest type" or their "strategic type". Cliquishness is around 70% (note that even at that high level, there are good chances of a scenario where the cliquishness does not lead to one-sidedness).
>
>
>What do others think is a reasonable scenario? I'd particularly like to hear from Clay. Clay, I know we're likely to disagree over the meaning of my VSE numbers once I have them ready; and, only human, we'll probably both tend to rationalize to make our points. I think that it will help ameliorate that disagreement if we undercut those future rationalizations by each precommitting to take our own chosen set of numbers seriously, even if those are two separate sets of numbers.
>
>
>Jameson
----
Election-Methods mailing list - see http://electorama.com/em for list info
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.electorama.com/pipermail/election-methods-electorama.com/attachments/20140224/a1ad4a9e/attachment-0004.htm>
More information about the Election-Methods
mailing list