SC4AI 2024: thank you and some thougths on AI for Social Choice
Arturo Macias
Dear all,
First, I want to say thank you to the organizing team for the workshop in Santiago de Compostela: it was a very interesting conference and a very interesting group of people. I want to write a few comments as a semi-outsider in the Social Choice Community.
Decision Theory Forum
Most of the people I interacted with in the Conference are not members of the Decision Theory Forum. It is a very large mailing list, owned by Itzhak Gilboa, and perhaps could be of interest for many of you.
Reinforcement Learning (RL) for Game Theory
In a few bilateral discussions, I always asked about the reverse issue to that of the workshop: what about Artificial Intelligence for Social Choice? Coincidentally, last week Marginal Revolution highlighted a paper on the use of RL techniques in dynamic Macroeconomics. The arguments advanced in the paper are even more relevant for the wide Game Theory community than for macroeconomics.
To some extent, after Alpha Zero the curse of dimensionality has been “tamed” and we can easily populate any “game” (=“virtual World” with known interaction rules) with superhuman optimizers and describe the “ergodic” distribution of the game outcomes (=its equilibrium). This implies a massive enlargement of our social modelling freedom.
Storable Votes and dynamic voting
Another interesting observation for me was that no presentation in the workshop was about “dynamic voting”. I became interested in voting systems after reading Weyl & Posner “Radical Markets”, where the authors propose “quadratic voting” as a mechanism to avoid “minority disenfranchisement” and allow the expression of preference intensity.
Quadratic voting makes sense when there is a stream of issues to be decided with different relevance for each voter. Then, agents want to “trade” votes from the issues they care less for votes in the issues that are more important for them.
A brilliant survey of the vote trading problem was published by Casella and Mace in their 2021 Annual Review “Does Vote Trading Improve Welfare?”[ungated version here]. Alessandra Casella proposed in 2003 the “storable votes” system, where instead of explicit bargaining she proposed that agents could be allowed to “store” some votes (withdrawing votes in low relevance elections to use them in the elections of more relevance to them). This simple idea led to a surprisingly powerful result in 2009 about implementability in repeated and related elections: In “Overcoming Incentive Constraints by Linking Decisions” it is shown that “the utility costs associated with incentive constraints become negligible when the decision problem is linked with a large number of independent copies of itself”. To some extent, this implies that in a dynamic setting, not only the Arrow impossibility vanishes, but even implementability problems become tractable.
RL for dynamic voting
I have recently made a contribution to this literature (see here my explanatory post in the Decision Theory Forum). Unfortunately, dynamic voting analysis is very affected by the curse of dimensionality, and my numeric exploration was very low dimensional. That is why I am looking for a coauthor with experience in reinforcement learning to explore the welfare properties of different dynamic voting mechanisms, including my own. Please feel free to send this to anybody you consider could be interested in this project.
Note: if you lack access to any cited paper, please let me know and I will send it to you.