Eshwar Ram Arunachaleswaran

What should swap-regret mean beyond normal-form games? We have a new paper (@natalie_collina , Mehryar Mohri, Yishay Mansour, Jon Schneider, Balu Sivan) tackling this question and providing a definitive answer! (thread) arxiv.org/abs/2502.20229

Based on some fun work connecting no-regret algorithms to pricing and collusion, with @natalie_collina, @Aaroth , Juba Ziani and my advisor Sam path Kannan

Suppose you and I both have different features about the same instance. Maybe I have CT scans and you have physician notes. We'd like to collaborate to make predictions that are more accurate than possible from either feature set alone, while only having to train on our own data.

We prove minimizing profile swap-regret is necessary & sufficient for non-manipulability and gets NR +PO. Bonus: if all agents minimize it, the dynamics can reach profiles that cannot be realized as Correlated Equilibria by traditional mediators—unlike normal-form games!

Our fix? Profile Swap-Regret, a further coarsening of polytope swap-regret, leveraging a geometric view of algorithms (link). Admits an efficient algorithm with O(√T) convergence! arxiv.org/abs/2402.09549

What should swap-regret mean beyond normal-form games? We have a new paper (@natalie_collina , Mehryar Mohri, Yishay Mansour, Jon Schneider, Balu Sivan) tackling this question and providing a definitive answer! (thread) arxiv.org/abs/2502.20229

@willccbb Very interesting! Just tried it out.Full chains of thought are amazing to read especially where the model makes some progress, can't quite get the answer and resorts to writing sanitized boilerplate text without speculating in the final output

@EshwarERA likely a relatively small model, mostly a demo of their reasoning chain tricks / teasing a bigger version coming soon

it was so close...

@willccbb Right. Which model is this? Pieces of the answer are correct traces of reasoning, but it's unable to chain them together

@EshwarERA yeah the internet word association btw no-swap + CE is way higher than what you get from the few dozen papers about no-swap stackelberg stuff

Now appearing in SODA; thanks to an eagle eyed reviewer, we've updated the title to "An Elementary Predictor Obtaining 2*Sqrt{T} + 1 Distance to Calibration." Fortunately it still fits in one line. (It turns out there are m+1 numbers in the set {0, 1/m, 2/m, ..., 1}, not m...)

@willccbb That's a cool explanation. Would digging into it require understanding how the underlying value model generalizes the estimates from training to the new states it encounters at test time? Thinking of mechanisms that would result in this imperfect correlation

yeah, let’s restrict to fully observable EFGs with clear win states you def need search during training to estimate a value function but it’s interesting that mtcs at test-time is still useful empirically — it wouldn’t be needed at all if you had a perfect value function i think you could also show that it wouldn’t help if value estimates were worst-case correlated (eg over a subtree) one plausible story is that test-time search helps you explore enough of the state space that value estimate errors are imperfectly correlated, so you get concentration of errors and basically a new value function with tighter confidence intervals

anyone have a good learning-theoretic explanation of why scaling search is more efficient than scaling models (e.g for alphago)? feels like there could maybe be a boosting-related story

@willccbb I'd be very interested in hearing any boosting like explanations. Are you perhaps hinting at search aggregating varying performance of an imperfect model over different parts of the context space?

The Sherman Act (1890) was written to prevent old fashioned price fixing: picture two men meeting in a smoky bar. It requires overt communication of intent to coordinate on prices. But pricing algorithms can automatically coordinate on high prices. What is algorithmic collusion?

What scientific overpublication is an inevitable symptom of the mass psychosis of a cabal of overachievers? argmin.net/p/youre-gonna-…

A quick thread on a short (3 page) paper, giving a simple algorithm that makes predictions guaranteeing 2*Sqrt{T} "Distance to calibration" against an adversary. The algorithm and proof are so simple I can describe it in thread. Joint with Eshwar, @natalie_collina, and Mirah:

A tweet thread about our recent paper on (Pareto) optimal learning algorithms for repeated games; i.e. how to learn to play against non-myopic opponents