Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

Full Swap Regret and Discretized Calibration

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that swap regret in structured d-dimensional games can be minimized at rate $\tilde{O}(T^{(d+1)/(d+3)})$, independent of the number of actions, and that the same machinery yields $\tilde{O}(T^{1/3})$ $\ell^2$-calibration…

desk verdict Real new benchmark and a clean structured-game reduction; the discretized calibration theorem needs a missing rounding lemma and corrected OGD coefficients before it is proven. read the letter →

arxiv 2502.09332 v1 pith:S6RMIFB7 submitted 2025-02-13 cs.LG cs.GT

Maxwell Fishelson , Robert Kleinberg , Princewill Okoroafor , Renato Paes Leme , Jon Schneider , Yifeng Teng This is my paper
classification cs.LGcs.GT MSC 68Q3268W2791A26
keywords swapregretfullstructuredgamesonlinecalibrationerrorconvexoptimizationcorrelatedequilibriumstronglylosses
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper's target is an efficient way to minimize swap regret, the regret measured against arbitrary remappings of past actions, when a player's many pure actions live as points in a low-dimensional Euclidean space and payoffs are bilinear in those points. In such d-dimensional structured games, it proves an algorithm can guarantee at most $\tilde{O}(T^{(d+1)/(d+3)})$ swap regret after $T$ rounds, with per-round time polynomial in $d$ and $T$ and no dependence on the number of actions. This matters because swap regret is the benchmark that drives play toward correlated equilibrium, and prior no-swap-regret methods either paid polynomially in the action count or exponentially in $1/\epsilon$. The same framework gives a forecasting algorithm with $\tilde{O}(T^{1/3})$ $\ell^2$-calibration error, and an $\epsilon$-discretized forecaster with $\tilde{O}(\max(\sqrt{\epsilon T},T^{1/3}))$ error, connecting a classical calibration problem to continuous swap regret.

What carries the argument

The load-bearing machinery is a continuous adaptation of the Blum–Mansour swap-regret template. One external-regret subroutine is run for each point of a finite discretization $\mathcal{K}_\epsilon$ of the convex action set; their mixed recommendations form the rows of a Markov chain, and the learner samples its stationary distribution, so swap regret decomposes into a sum of scaled external regrets. To make this work on convex domains, the paper combines geometric discretizations, an $\epsilon$-net, an $\epsilon$-triangulation, or a polytope approximation, whose rounding error is controlled by Lipschitzness or smoothness; a rounding procedure $H:\mathcal{K}\to\Delta(\mathcal{K}_\epsilon)$; scaled online gradient descent as the external-regret subroutine with learning-rate schedules tuned for strongly convex and for $(\alpha,\epsilon)$-nearly-strongly-convex losses; and a piecewise-linearization of discretized losses that converts discretized calibration into a nearly-strongly-convex problem. The reduction from structured games to linear full swap regret is the bridge that turns these continuous-regret algorithms into algorithms for normal-form games with exponentially many pure strategies.

What would settle it

Implement the algorithm from Theorem 21 on one-dimensional discretized calibration with grid spacing $\epsilon = T^{-1/4}$, adversarial outcomes $b_t \in \{0,1\}$ chosen as $b_t = 1$ whenever the algorithm's forecast is below $1/2$ and $0$ otherwise, and measure the realized discretized calibration error over $T = 10^6$ rounds; if the error grows faster than a constant times $T^{3/8}\log T$, the dominant term of $\max(\sqrt{\epsilon T},T^{1/3})$, the claimed bound is contradicted.

Watch

Extended reading notes

Core claim

The central discovery is that full swap regret, the natural continuous version of swap regret defined against every map $\varphi:\mathcal{K}\to\mathcal{K}$ from a convex action set to itself, can be minimized at rates that depend on the dimension and on the loss geometry rather than on the number of actions. The paper establishes a reduction showing that swap regret in any d-dimensional structured game is bounded by full swap regret against linear losses on the embedding hull $\mathcal{K}$, and then proves a family of full-swap-regret bounds: $\tilde{O}(T^{(d+1)/(d+2)})$ for losses with no further assumption, improved to $\tilde{O}(T^{(d+1)/(d+3)})$ for linear or concave losses, $\tilde{O}(T^{(d+2)/(d+4)})$ for smooth losses, and up to $\tilde{O}(T^{d/(d+2)})$ when losses are both strongly convex and smooth. When $d=1$, these are instantiated to calibration by identifying $\ell^2$-calibration error with swap regret against the quadratic losses $(x-b_t)^2$, yielding an $\tilde{O}(T^{1/3})$ bound; for predictions restricted to multiples of $\epsilon$, piecewise linearization plus a new nearly-strongly-convex analysis yields $\tilde{O}(\max(\sqrt{\epsilon T},T^{1/3}))$.

Load-bearing premise

The efficient end-to-end result assumes the Learner observes the adversary's embedding of its mixed action each round and has a poly(d) oracle that converts any point of the embedding hull into a mixed strategy over the original actions; without those, the claimed per-round poly(d,T) runtime and the (1/epsilon)^{O(d)} correlated-equilibrium computation fail, though the regret bound may degrade by only an O(n) factor.

Editorial extensions

If this is right

  • For any constant embedding dimension $d$, per-round swap regret becomes $\tilde{O}(T^{(d+1)/(d+3)})$ independent of the number of pure actions, so games with exponentially many strategies remain tractable.
  • Two players running the algorithms converge to an $\epsilon$-correlated equilibrium in $T = \tilde{\Omega}((1/\epsilon)^{(d+3)/2})$ rounds, and an $\epsilon$-correlated equilibrium can be computed in time $(1/\epsilon)^{O(d)}$.
  • Online $\ell^2$-calibration error drops to $\tilde{O}(T^{1/3})$, improving on the $\Omega(\sqrt{T})$ rate of naively applying discrete swap-regret algorithms to a fine prediction grid.
  • When forecasts must be multiples of $\epsilon$, the bound $\tilde{O}(\max(\sqrt{\epsilon T},T^{1/3}))$ beats both rounding the continuous forecaster and running Blum–Mansour directly on the grid, across a wide intermediate range of $\epsilon$.
  • Full swap regret rates improve with loss curvature, moving from exponent $(d+1)/(d+2)$ with no assumptions to $d/(d+2)$ when losses are both strongly convex and smooth.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The two-phase learning-rate schedule for $(\alpha,\epsilon)$-nearly-strongly-convex functions is developed for external regret, and it should transfer to any online convex optimization setting where the loss is strongly convex at scales above $\epsilon$ but nearly linear at small scales, potentially improving regret for discretized decision problems beyond calibration.
  • The lossless rounding used for discretized calibration is special to one dimension, and the paper explicitly leaves the higher-dimensional analogue open; finding one would extend the same template to multi-dimensional calibration and structured games.
  • Because any normal-form game with $\min(n,n')$ actions is itself a $\min(n,n')$-dimensional structured game, the framework offers a unified view of when low-dimensional structure, rather than action count, is the right complexity parameter for swap regret.
  • The exact equivalence between $\ell^2$-calibration error and full swap regret suggests that other proper scoring rules whose losses are strongly convex and smooth could inherit similar $T^{1/3}$-type calibration guarantees from the same reduction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces a new online learning benchmark, full swap regret over a convex action set K in d dimensions, and shows that it captures swap regret in structured normal-form games via a reduction (Lemma 11). The main algorithmic contribution is a Blum-Mansour style template for convex action sets, instantiated with different discretizations and external-regret subroutines; the resulting rates are summarized in Table 5 (Theorem 15), ranging from O~(T^{(d+1)/(d+2)}) to O~(T^{(d)/(d+2)}) depending on loss assumptions. The paper then applies the framework to online calibration: Theorem 4 gives O~(T^{1/3}) l2-calibration error, and Theorem 5 claims O~(max(T^{1/3}, sqrt(epsilon T))) discretized-calibration error. The proof of the latter rests on a new external-regret algorithm for nearly-strongly-convex losses (Theorem 20) and a discretized swap-regret bound (Theorem 21).

Significance. If the results are correct, the paper makes a substantial contribution: it gives the first dimension-dependent swap-regret rates for structured games that are independent of the number of actions, and it connects full swap regret to calibration in a clean way. The proposed full swap regret notion, the reduction from structured games (Lemma 11), and the nearly-strongly-convex OGD variant are all potentially useful tools. The claimed O(T^{1/3}) l2-calibration bound and the interpolation for discretized calibration are also significant improvements over the naive baselines discussed in the paper. However, the central discretized-calibration theorem (Theorem 5) and the theorem it depends on (Theorem 21) are not proved as stated; the proof gaps are load-bearing and must be repaired before the main claims can be accepted.

major comments (3)
  1. [Appendix D.4 / Theorem 5] The proof of Theorem 5 says the result 'follows directly from Theorem 21', but Theorem 21 gives O(L sqrt(epsilon T) + L^2/(alpha epsilon) log T). For epsilon < T^{-1/3}, the second term epsilon^{-1} log T dominates T^{1/3}, so the claimed O~(max(T^{1/3}, sqrt(epsilon T))) bound does not follow. A missing rounding argument (e.g., a lemma bounding the grid-rounding distortion of Cal_epsilon by O(R + epsilon^2 T)) is exactly what would close this gap, but no such lemma appears. This is a central claim of the paper, so the gap must be fixed or the theorem restated with the correct regime.
  2. [Appendix D.7 / proof of Theorem 21] The step labeled 'Lemma 10 gives' contains an arithmetic error. Summing the per-subroutine bound 2 sqrt(2 epsilon L) sqrt(G_s) over the |K_epsilon| = 1/epsilon states gives, by Cauchy-Schwarz, at most 2 sqrt(2 epsilon L) sqrt(T/epsilon) = O(L sqrt(T)), not O(L sqrt(epsilon T)) as printed. The intermediate expression inserts an extra factor of epsilon and silently drops a factor 1/sqrt(epsilon). Consequently the sqrt(epsilon T) term in Theorem 21 is unsupported as written. The same issue propagates to Theorem 6 and to the proof of Theorem 5.
  3. [Appendix D.7 / Eq. (7) in proof of Theorem 20] The display '(7)' uses the coefficient alpha/2 in the indicator term, but the (alpha, epsilon)-nearly-strong-convexity definition (Definition 18) gives only alpha/8 on the event ||x_t - x*|| >= 2 epsilon, since (r - epsilon)_+^2 >= r^2/4 implies alpha/2 * r^2/4 = alpha r^2/8. This constant-factor slip is in a load-bearing part of the proof of Theorem 20 and should be corrected; it does not change the asymptotic rate but it does invalidate the stated constants and the derivation of the interpolation threshold.
minor comments (5)
  1. [Appendix D.4] The proof of Theorem 5 says 'we can plug into Theorem 5'; this should be 'Theorem 21'.
  2. [Appendix D.4] The sentence 'For the regime where epsilon = o(T^{1/3})' should be 'for the regime where epsilon^{-1} = o(T^{1/3})' (equivalently epsilon = omega(T^{-1/3})), otherwise the comparison to Theorem 4 is inconsistent.
  3. [Theorem 4 proof, Appendix D.3] The proof refers to 'Case 3 of Theorem 15', but Table 5 has five rows and the calibration argument uses the strongly-convex and smooth row; the reference should be to the corresponding row, not 'Case 3'.
  4. [Lemma 26 proof, Appendix D.2] The sentence 'as shown in Theorem 26' should refer to Lemma 26.
  5. [Appendix D.7] The proof of Theorem 20 contains a stray LaTeX artifact '/BD' in the display before the indicator; this should be removed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the derivation chain is self-contained, with the only flagged concern being a correctness gap rather than a circular reduction.

full rationale

Walking the derivation chain: Theorem 1 is obtained from the full-swap-regret bound for linear/concave losses in Theorem 15 via the reduction in Lemma 11; the reduction compares the game swap regret to FullSwapReg against the sequence of linear losses x -> -<x, pi'(q_t)> and is proven directly, so the structured-game claim does not presuppose itself. Theorem 15 is an instance of the Blum-Mansour template (Theorem 12), with discretization and rounding lemmas (Lemmas 13 and 14) bounding the approximation error and with external-regret subroutines (Lemma 24 for strongly convex losses, Theorem 20 for nearly strongly convex losses) whose proofs are self-contained or cited to standard OCO references. Theorem 4 follows by instantiating Theorem 15 with K=[0,1] and the 2-strongly-convex, 2-smooth squared losses, using Lemma 26 to identify calibration error with full swap regret; that equivalence is proven rather than definitional. Theorem 21 similarly uses the lossless rounding of Lemma 17 and the nearly-strongly-convex external regret bound of Theorem 20. No parameter is fitted to the target calibration bound, and the only comparisons to prior work (Dagan et al., Peng and Rubinstein, etc.) are contextual rather than load-bearing. The author-overlapping citations appear in related-work and comparison passages, but no central equation is justified by those citations. The one flagged issue is a correctness/proof gap, not circularity: Appendix D.4 claims Theorem 5 follows directly from Theorem 21, but the stated O(sqrt(epsilon T) + epsilon^{-1} log T) bound is not shown to imply O(max(T^{1/3}, sqrt(epsilon T))) for epsilon < T^{-1/3}, and the proof of Theorem 21 in Appendix D.7 appears to contain a coefficient slip. These are gaps in the proof as printed, not reductions of a prediction to its own input, so they do not raise the circularity score.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The central claims rest on standard geometric, convex-optimization, and regret-decomposition facts, plus the domain assumptions of bounded, Lipschitz, strongly-convex/smooth losses and the availability of a convex decomposition oracle. No free parameters are fitted to data; the discretization scale ǫ is chosen analytically to balance regret terms. No new physical or metaphysical entities are introduced.

assumptions (5)
  • standard math Existence of ǫ-nets and ǫ-triangulations of size O(ǫ^{-d}) for convex sets in the unit ball (Lemma 8, Lemma 10, relying on Bronshteyn-Ivanov).
    Used for the discretization in Algorithms 1 and 2; proof of Lemma 10 is given in Appendix A.
  • domain assumption Loss functions are L-Lipschitz and, in the respective rows, α-strongly-convex and/or β-smooth.
    Defines the problem classes in Theorem 15; the calibration losses satisfy these with L=2, α=2, β=2.
  • domain assumption In structured games, the Learner has access to an efficient convex decomposition oracle π^{-1}(x) and observes the Adversary's embedding y_t each round.
    Needed in Lemma 11 to convert full-swap-regret algorithms into poly(d,T) per-iteration algorithms for the original game; without it, runtime becomes O(n,n').
  • standard math Standard OGD and multiplicative weights regret bounds, including the scaled-regret versions in Lemma 24.
    Used as subroutines inside the Blum-Mansour template; proofs are standard and included in Appendix C.
  • standard math Blum-Mansour Markov-chain decomposition of swap regret into external regrets.
    The template of Theorem 12; proof in Appendix B.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Full Swap Regret and Discretized Calibration." pith.science (2026). https://pith.science/paper/S6RMIFB7

@misc{pith2026250209332,
  author       = {Pith},
  title        = {Pith review of: Full Swap Regret and Discretized Calibration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S6RMIFB7}},
  note         = {Machine review of arXiv:2502.09332}
}
abstract

We study the problem of minimizing swap regret in structured normal-form games. Players have a very large (potentially infinite) number of pure actions, but each action has an embedding into $d$-dimensional space and payoffs are given by bilinear functions of these embeddings. We provide an efficient learning algorithm for this setting that incurs at most $\tilde{O}(T^{(d+1)/(d+3)})$ swap regret after $T$ rounds. To achieve this, we introduce a new online learning problem we call \emph{full swap regret minimization}. In this problem, a learner repeatedly takes a (randomized) action in a bounded convex $d$-dimensional action set $\mathcal{K}$ and then receives a loss from the adversary, with the goal of minimizing their regret with respect to the \emph{worst-case} swap function mapping $\mathcal{K}$ to $\mathcal{K}$. For varied assumptions about the convexity and smoothness of the loss functions, we design algorithms with full swap regret bounds ranging from $O(T^{d/(d+2)})$ to $O(T^{(d+1)/(d+2)})$. Finally, we apply these tools to the problem of online forecasting to minimize calibration error, showing that several notions of calibration can be viewed as specific instances of full swap regret. In particular, we design efficient algorithms for online forecasting that guarantee at most $O(T^{1/3})$ $\ell_2$-calibration error and $O(\max(\sqrt{\epsilon T}, T^{1/3}))$ \emph{discretized-calibration} error (when the forecaster is restricted to predicting multiples of $\epsilon$).

Figures

Figures reproduced from arXiv: 2502.09332 by the authors.

Figure 1
Figure 1. The x-axis shows the discretization parameter ǫ and the y-axis the calibration loss Cal for three different algorithms: dashed blue (regular swap regret on 1/ǫ actions), dashed black (algorithm in Theorem 4 + rounding) and red (algorithm in Theorem 5). convex, then there exists an algorithm that incurs at most O( √ ǫT +ǫ −1 log T) swap regret (against the best map φ : K → K). The main idea behind the proof of Theore… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. High-Dimensional Calibration from Swap Regret

    cs.LG 2025-05 conditional novelty 7.0 of 10

    TreeCal achieves epsilon-calibration over arbitrary convex sets and norms in (diam/eps)^{O(rho/eps^2)} rounds, and a new lower bound shows exp(poly(1/eps)) rounds are necessary for l1-calibration on the simplex.

  2. Persuasive Prediction via Decision Calibration

    cs.GT 2025-05 reject novelty 6.0 of 10

    A data-driven sender can learn a near-optimal decision-calibrated predictor without knowing the prior, but the proof as written has a critical Lagrangian error and the Bayesian benchmark is restricted by construction.

Reference graph

Works this paper leans on

38 extracted references · 23 canonical work pages · cited by 2 Pith papers

  1. [1]

    Efficient phi-regret minimization in extensive-form games via online mirror descent

    Yu Bai, Chi Jin, Song Mei, Ziang Song, and Tiancheng Yu. Efficient phi-regret minimization in extensive-form games via online mirror descent. Advances in Neural Information Processing Systems, 35: 0 22313--22325, 2022

    work page 2022

  2. [2]

    From external to internal regret

    Avrim Blum and Yishay Mansour. From external to internal regret. Journal of Machine Learning Research, 8 0 (6), 2007

    work page 2007

  3. [3]

    Selling to a no-regret buyer

    Mark Braverman, Jieming Mao, Jon Schneider, and Matt Weinberg. Selling to a no-regret buyer. In Proceedings of the 2018 ACM Conference on Economics and Computation, pages 523--538, 2018

    work page 2018

  4. [4]

    Verification of forecasts expressed in terms of probability

    Glenn W Brier. Verification of forecasts expressed in terms of probability. Monthly weather review, 78 0 (1): 0 1--3, 1950

    1950

  5. [5]

    The approximation of convex sets by polyhedra

    Efim M Bronshteyn and LD Ivanov. The approximation of convex sets by polyhedra. Siberian Mathematical Journal, 16 0 (5): 0 852--853, 1975

    work page 1975

  6. [6]

    Superhuman ai for heads-up no-limit poker: Libratus beats top professionals

    Noam Brown and Tuomas Sandholm. Superhuman ai for heads-up no-limit poker: Libratus beats top professionals. Science, 359 0 (6374): 0 418--424, 2018

    work page 2018

  7. [7]

    Is learning in games good for the learners? Advances in Neural Information Processing Systems, 36, 2024

    William Brown, Jon Schneider, and Kiran Vodrahalli. Is learning in games good for the learners? Advances in Neural Information Processing Systems, 36, 2024

    2024

  8. [8]

    Selling to multiple no-regret buyers

    Linda Cai, S Matthew Weinberg, Evan Wildenhain, and Shirley Zhang. Selling to multiple no-regret buyers. In International Conference on Web and Internet Economics, pages 113--129. Springer, 2023

    work page 2023

Show all 38 references
  1. [9]

    Mechanisms for a no-regret agent: Beyond the common prior

    Modibo K Camara, Jason D Hartline, and Aleck Johnsen. Mechanisms for a no-regret agent: Beyond the common prior. In 2020 ieee 61st annual symposium on foundations of computer science (focs), pages 259--270. IEEE, 2020

    2020

  2. [10]

    Decentralized no-regret learning algorithms for extensive-form correlated equilibria

    Andrea Celli, Alberto Marchesi, Gabriele Farina, Nicola Gatti, et al. Decentralized no-regret learning algorithms for extensive-form correlated equilibria. In IJCAI, pages 4755--4759, 2021

    2021

  3. [11]

    Prediction, learning, and games

    Nicolo Cesa-Bianchi and G \'a bor Lugosi. Prediction, learning, and games. Cambridge university press, 2006

    2006

  4. [12]

    Efficient learning in polyhedral games via best-response oracles

    Darshan Chakrabarti, Gabriele Farina, and Christian Kroer. Efficient learning in polyhedral games via best-response oracles. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 9564--9572, 2024

    2024

  5. [13]

    From external to swap regret 2.0: An efficient reduction and oblivious adversary for large action spaces

    Yuval Dagan, Constantinos Daskalakis, Maxwell Fishelson, and Noah Golowich. From external to swap regret 2.0: An efficient reduction and oblivious adversary for large action spaces. arXiv preprint arXiv:2310.19786, 2023

    2023 arXiv

  6. [14]

    Breaking the t^ 2/3 barrier for sequential calibration, 2024

    Yuval Dagan, Constantinos Daskalakis, Maxwell Fishelson, Noah Golowich, Robert Kleinberg, and Princewill Okoroafor. Breaking the t^ 2/3 barrier for sequential calibration, 2024. URL https://arxiv.org/abs/2406.13668

    2024 arXiv

  7. [15]

    Efficient learning and computation of linear correlated equilibrium in general convex games

    Constantinos Daskalakis, Gabriele Farina, Maxwell Fishelson, Charilaos Pipis, and Jon Schneider. Efficient learning and computation of linear correlated equilibrium in general convex games. arXiv preprint arXiv:2412.20291, 2024

    2024 arXiv

  8. [16]

    Prior-free dynamic auctions with low regret buyers

    Yuan Deng, Jon Schneider, and Balasubramanian Sivan. Prior-free dynamic auctions with low regret buyers. Advances in Neural Information Processing Systems, 32, 2019 a

    2019

  9. [17]

    Strategizing against no-regret learners

    Yuan Deng, Jon Schneider, and Balasubramanian Sivan. Strategizing against no-regret learners. Advances in neural information processing systems, 32, 2019 b

    2019

  10. [18]

    Polynomial-time linear-swap regret minimization in imperfect-information sequential games

    Gabriele Farina and Charilaos Pipis. Polynomial-time linear-swap regret minimization in imperfect-information sequential games. Advances in Neural Information Processing Systems, 36, 2024

    2024

  11. [19]

    Calibrated learning and correlated equilibrium

    Dean P Foster and Rakesh V Vohra. Calibrated learning and correlated equilibrium. Games and Economic Behavior, 21 0 (1-2): 0 40--55, 1997

    1997

  12. [20]

    Asymptotic calibration

    Dean P Foster and Rakesh V Vohra. Asymptotic calibration. Biometrika, 85 0 (2): 0 379--390, 1998

    1998

  13. [21]

    Game theory, on-line prediction and boosting

    Yoav Freund and Robert E Schapire. Game theory, on-line prediction and boosting. In Proceedings of the ninth annual conference on Computational learning theory, pages 325--332, 1996

    1996

  14. [22]

    Generative adversarial networks

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63 0 (11): 0 139--144, 2020

    2020

  15. [23]

    Omnipredictors

    Parikshit Gopalan, Adam Tauman Kalai, Omer Reingold, Vatsal Sharan, and Udi Wieder. Omnipredictors. In 13th Innovations in Theoretical Computer Science Conference (ITCS 2022). Schloss Dagstuhl-Leibniz-Zentrum f \"u r Informatik, 2022

    2022

  16. [24]

    Gordon, Amy Greenwald, and Casey Marks

    Geoffrey J. Gordon, Amy Greenwald, and Casey Marks. No-regret learning in convex games. In Proceedings of ICML , volume 307, pages 360--367. ACM , 2008

    2008

  17. [25]

    Geometric algorithms and combinatorial optimization, volume 2

    Martin Gr \"o tschel, L \'a szl \'o Lov \'a sz, and Alexander Schrijver. Geometric algorithms and combinatorial optimization, volume 2. Springer Science & Business Media, 2012

    2012

  18. [26]

    Calibrated stackelberg games: Learning optimal commitments against calibrated agents

    Nika Haghtalab, Chara Podimata, and Kunhe Yang. Calibrated stackelberg games: Learning optimal commitments against calibrated agents. Advances in Neural Information Processing Systems, 36, 2024

    2024

  19. [27]

    Introduction to online convex optimization

    Elad Hazan. Introduction to online convex optimization. MIT Press, 2022

    2022

  20. [28]

    Calibration error for decision making, 2024

    Lunjia Hu and Yifan Wu. Calibration error for decision making, 2024. URL https://arxiv.org/abs/2404.13503

    2024 arXiv

  21. [29]

    U-calibration: Forecasting for an unknown agent

    Bobby Kleinberg, Renato Paes Leme, Jon Schneider, and Yifeng Teng. U-calibration: Forecasting for an unknown agent. In The Thirty Sixth Annual Conference on Learning Theory, pages 5143--5145. PMLR, 2023

    2023

  22. [30]

    Strategizing against learners in bayesian games

    Yishay Mansour, Mehryar Mohri, Jon Schneider, and Balasubramanian Sivan. Strategizing against learners in bayesian games. In Conference on Learning Theory, pages 5221--5252. PMLR, 2022

    2022

  23. [31]

    Scalar and vector partitions of the probability score: Part i

    Allan H Murphy. Scalar and vector partitions of the probability score: Part i. two-state situation. Journal of Applied Meteorology and Climatology, 11 0 (2): 0 273--282, 1972

    1972

  24. [32]

    A new vector partition of the probability score

    Allan H Murphy. A new vector partition of the probability score. Journal of Applied Meteorology and Climatology, 12 0 (4): 0 595--600, 1973

    1973

  25. [33]

    Fast swap regret minimization and applications to approximate correlated equilibria

    Binghui Peng and Aviad Rubinstein. Fast swap regret minimization and applications to approximate correlated equilibria. arXiv preprint arXiv:2310.19647, 2023

    2023 arXiv

  26. [34]

    Stronger calibration lower bounds via sidestepping

    Mingda Qiao and Gregory Valiant. Stronger calibration lower bounds via sidestepping. In Proceedings of the 53rd Annual ACM SIGACT Symposium on Theory of Computing, pages 456--466, 2021

    2021

  27. [35]

    Forecasting for swap regret for all downstream agents

    Aaron Roth and Mirah Shi. Forecasting for swap regret for all downstream agents. In Proceedings of the 25th ACM Conference on Economics and Computation, pages 466--488, 2024

    2024

  28. [36]

    Mastering the game of go without human knowledge

    David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go without human knowledge. nature, 550 0 (7676): 0 354--359, 2017

    2017

  29. [37]

    High-dimensional probability: An introduction with applications in data science, volume 47

    Roman Vershynin. High-dimensional probability: An introduction with applications in data science, volume 47. Cambridge university press, 2018

    2018

  30. [38]

    Efficient phi-regret minimization with low-degree swap deviations in extensive-form games

    Brian Hu Zhang, Ioannis Anagnostides, Gabriele Farina, and Tuomas Sandholm. Efficient phi-regret minimization with low-degree swap deviations in extensive-form games. arXiv preprint arXiv:2402.09670, 2024

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.