Idea for simulations
-
One thing we could do is take all 440 RCV elections that have happened and for which ballot data is available, and try to convert each of them to a spatial model, such as this one:
Ideology is represented in 2 dimensions as X and Y, while a third factor, "universal appeal" is shown by the size of the candidate. There is no need to restrict ideology to 2 dimensions, though.... the only reason I do it is to keep it easy for humans to look at it on a 2d screen and reason about it.
Once we have spatial models, we could perturb various things (such as the positions of the candidates, where the voters are, etc) and regenerate ballots, for various sorts of methods and with appropriate strategy.
It's not all that easy to go backwards from ballots to a spatial model, but it isn't that hard. It's similar to the logic used for recommendation systems. It's been a good 15 years since I used that sort of logic in the Netflix contest and did pretty well.
In theory, someone could train an AI model on election data and use that to make simulated elections scenarios. That's beyond my own capabilities, but I know a lot of people who could probably do it quickly if they had the data in a convenient format.