Pith. sign in

REVIEW 4 major objections 5 minor 34 references

Scalable Policy Optimization for Networked Multi-Agent Reinforcement Learning with Continuous State-Action Spaces

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read Under exponential spatial decay and an explicit feature-excitation condition, CDCPG drives averaged per-agent stationarity to within any excess of a structural floor using Õ(ε⁻²) shared-oracle samples.

desk verdict Conditional result with real new machinery: an O~(1/ε²) guarantee that is honest about resting on an unverified TD-stability premise. read the letter →

arxiv 2607.18554 v1 pith:HFN4DNYS submitted 2026-07-20 cs.MA cs.LG

classification cs.MAcs.LG MSC 68T0568W4090C40
keywords networkedmulti-agentreinforcementlearningcontinuousstate-actionspacespolicygradientrandomFourierfeaturestemporal-differencestabilityexponentialspatialdecaysharedgenerativeoraclesamplecomplexity
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

CDCPG is a cooperative multi-agent reinforcement learning algorithm for networked Markov decision processes with continuous state and action spaces. Each agent maintains a local policy and a localized spectral critic built from random Fourier features, and updates using a policy gradient estimator that aggregates neighboring critics. The paper's central claim is that, under exponential spatial decay of interactions, a symmetric feature-excitation condition, and smoothness, the averaged per-agent stationarity measure can be driven to within any excess above an explicit structural floor using Õ(ε⁻²) shared-oracle samples, with per-agent computation and communication governed by neighborhood size rather than network size. If correct, this would be the first continuous state-action analogue of the exponential-decay scalability framework previously established for discrete networked MDPs, with the temporal-difference stability constant made explicit and monitorable.

What carries the argument

The central object is the conditional-expectation truncated Q-function, defined as the expected full Q-function given a local state-action window, which gives a well-posed localized Bellman operator. The analysis also relies on normalized random Fourier features with explicit global L1 approximation control, the population TD matrix M_θ_i = H_θ_i - γ C_θ_i, and the unconditional bound σ_min(M_θ_i) ≥ (1-√γ) λ_min(H_θ_i). This bound reduces temporal-difference stability to a symmetric feature-excitation condition, which is monitored through a matrix-Bernstein diagnostic.

What would settle it

Compute, for a continuous networked linear-Gaussian system with normalized random Fourier features, the population TD matrix under discounted-occupancy sampling and evaluate the minimum over agents and parameters of σ_min(M_θ) as the feature dimension grows. If this infimum is zero, or if the empirical λ_min of the symmetric feature covariance collapses toward zero, the excitation premise fails and the Õ(ε⁻²) excess-to-floor guarantee degenerates.

Watch

Extended reading notes

Core claim

The paper establishes a finite-sample guarantee for CDCPG: under a discounted-occupancy sampling model, if the population temporal-difference matrix M_θ_i satisfies a uniform lower bound on its smallest singular value (equivalently, the symmetric feature covariance H_θ_i has strictly positive smallest eigenvalue), and if the value functions decay exponentially with graph distance, then the averaged per-agent projected-gradient stationarity is bounded by ε plus a four-component structural floor, and reaching excess ε costs Õ(ε⁻²) shared-oracle samples. The proof also exposes a dimensional obstruction stating that no dimension-free lower bound on the TD matrix exists, and reduces stability to

Load-bearing premise

The entire sample-complexity guarantee depends on the unverified premise that the population temporal-difference matrix has a strictly positive lower bound on its smallest singular value, and the paper proves only an upper bound while explicitly leaving positive-excitation verification for concrete continuous model classes open.

Editorial extensions

If this is right

  • The guarantee matches the smooth nonconvex first-order rate: Õ(ε⁻²) samples to any excess above the floor, with the optimization term contributing the standard 1/ε dependence.
  • Per-agent computation and communication depend only on neighborhood sizes D_κ and D_{κ_c}, not on the number of agents n, enabling scalability to large networks.
  • The structural floor is explicit: it separates within-bin fluctuation, truncation residual, random-feature residual, and graph-tail aggregation, guiding feature-dimension and locality-radius choices.
  • The matrix-Bernstein diagnostic gives a post-hoc lower-confidence bound on the TD-stability constant along realized iterates, making the conditioning premise testable in practice.

Reading between the lines

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

  • A concrete population-level computation of λ_min(H_θ) on a linear-Gaussian networked model would either verify the excitation bound h_M(m)>0 or expose its failure; the paper leaves this open.
  • The adaptive-locality rule suggests that the optimal radius grows logarithmically in 1/ε on path-like graphs, but the feature-scale constant grows exponentially in the neighborhood dimension, so practical implementations may need to keep m fixed.
  • The stopped-process inequality of Theorem 7 could be converted into an online stopping rule that certifies the realized trajectory, closing the gap between prospective and post-hoc certification.
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

4 major / 5 minor

Summary. The paper proposes CDCPG, a distributed actor-critic algorithm for networked MDPs with continuous state and action spaces. Each agent maintains a localized policy and a least-squares temporal-difference critic built from spectral random Fourier features, with truncation defined through conditional expectations over a graph neighborhood. The main theoretical result (Theorem 6) states that, under Assumption Sets A–D — including an exponential spatial decay property, smoothness of the objective, and a temporal-difference stability condition σ_min(M_i^θ) ≥ μ_M(m) > 0 (Assumption 4) — CDCPG drives an averaged per-agent projected-gradient stationarity measure to within any excess ε of an explicit structural floor using Õ(ε^{-2}) shared-oracle samples. A trajectory-conditional variant (Theorem 7) replaces uniform stability with a pathwise premise and a stopped-process inequality. The paper also proves a dimensional obstruction (Proposition 2), reduces TD stability to a symmetric excitation condition (Lemma 3), provides a matrix-Bernstein false-certification diagnostic (Proposition 3), and gives an adaptive-locality rule (Proposition 4). Numerical experiments on a networked linear-quadratic benchmark study locality, feature dimension, decay, and regularization.

Significance. If the main theorem holds as stated, this is a substantive step: it would be the first continuous state-action analogue of the Qu/Lin/Ren exponential-decay framework with an explicit four-component error decomposition, and the paper is unusually transparent about its caveats. The proof scaffolding is elaborate and largely self-consistent: the conditional-expectation truncation fixes a real gap in naive Bellman truncation, Lemma 3 derives the cross-term bound from the occupancy identity rather than postulating it, and the paper explicitly disclaims minimax optimality and notes that rates across rows in Table I are not directly comparable. There is no fitted-parameter circularity: the theorem constants are not tuned to match experiments, and the structural floor is stated explicitly. However, the central scalability claim rests on a premise that is acknowledged to be unverified for any concrete continuous model class (Section V, after Lemma 3), and the experiments do not exercise the analyzed RFF–LSTD critic in the main comparison. The conditional theorems may be correct, but the paper does not instantiate its own core assumptions, which limits the significance of the Õ(ε^{-2}) guar

major comments (4)
  1. [Section V, Assumption 4 and Lemma 3; Section VI, Theorem 6] The central complexity bound (Theorem 6, Eq. (47)) depends polynomially on 1/μ_M through M_s^* = Õ(L^6 W^{*2} G^2 D_{κ_c}^2/(μ_M^2(1-γ)^2 ε)) and through the structural floor constants C' and C_bin (Eq. (33)). But the only general result about μ_M is the upper bound μ_M(m) ≤ (1+γ)L^2/√(m+1) (Proposition 2), and the paper explicitly states, after Lemma 3, that “verifying the excitation bound h_M(m)>0 for a concrete continuous networked model class is left open.” Lemma 3 is a reduction, not a verification: it shows σ_min(M_i^θ) ≥ (1-√γ)λ_min(H_i^θ), but it does not establish λ_min(H_i^θ) ≥ h_M(m)>0. Consequently, no concrete model class is shown to satisfy Assumption 4, and the eO(ε^{-2}) guarantee is conditional on an uninstantiated premise. This is load-bearing: if h_M=0 or μ_M decays faster than m^{-1/2}, the floor terms blow up (Remark 3), and Theorem 6 becomes vacuous for that class.
  2. [Section VII, subsections A and C] The main numerical comparison (Figs. 1–3, Table III) uses a structured quadratic-basis critic tailored to the LQR benchmark, not the augmented RFF–LSTD critic analyzed by Theorems 3–6. The paper is honest about this, but it means the experiments do not validate the analyzed algorithm's central stability mechanism; the m-sweep (Fig. 4) does use the RFF critic, yet the paper itself notes (Section VII.C) that the conditioning σ_min(M_i^(k)) drifts with m and that the trend is confounded with truncation and clipping mismatch. The empirical conditioning proxy μ̂_M ≈ 0.41 is reported for the structured critic only, and the paper states it “neither certifies (C′) nor estimates the population conditioning of the analyzed RFF critic.” For a paper whose central claim is a sample-complexity guarantee predicated on μ_M, the lack of any reported conditioning data for the analyzed RFF critic across m
  3. [Section V, Proposition 3 and Theorem 7] The trajectory-conditional companion (Theorem 7) is presented as a strong result, but its high-probability content (statements (c)–(d)) requires Assumption 5 as a prospective premise. Proposition 3's diagnostic only bounds the joint false-certification probability P(T_cert ∩ {inf σ_min < r_M}); it does not control P(inf σ_min < r_M | T_cert), and the paper explicitly notes that a test that never passes satisfies (52) vacuously. Moreover, as stated in Section V, the pilot-run calibration “does not certify the main trajectory.” Thus the diagnostic is post-hoc and cannot be used to select μ_traj or M_s before the run. This is not a logical error — the paper labels (C′) as a premise — but it means Theorem 7 does not provide a prospective guarantee unless Assumption 5 is independently verified. The distinction is already in the text, but the abstract and contributions section present Theorem
  4. [Section VI, Remark 3 and Theorem 6] The eO(1/ε^2) rate is stated 'to within a structural floor' (Eq. (46)), but Remark 3 shows that the certified generic within-bin floor grows at least linearly in m under the best-case conditioning scaling μ_M = Θ(m^{-1/2}): (1+C')^2 = Θ(m), C_bin^2 = Θ(m), and W^{*2} = O(m). Since m is treated as a problem primitive fixed before ε → 0, the floor can easily exceed a target τ unless m is chosen carefully; yet no concrete choice of (κ,m) is shown to make ϵ_floor(κ,m) below a given τ while preserving the assumptions. Proposition 4 is an oracle diagnostic, not a rate improvement, and the paper explicitly says the inflation can violate condition (i) for fixed m. This does not invalidate the conditional theorem, but it means the practical scalability claim — that CDCPG achieves near-ε stationarity with Õ(1/ε^2) samples — is not demonstrated for any concrete parameter regime. The authors should
minor comments (5)
  1. [Abstract and Section I.B] The abstract states the sample complexity as Õ(ε^{-2}) without immediately qualifying 'to a structural floor' and 'under unverified structural premises.' Consider adding the qualifiers in the abstract itself to avoid overstatement.
  2. [Section II.C, Assumption 3 (D1)] The shared-sample oracle model is described as 'generative' with samples from d^π, but later the paper notes that realizing each call by geometric rollouts costs an extra 1/(1-γ) factor. This is acknowledged in passing; please state the environment-step complexity explicitly in the main complexity theorem.
  3. [Section V, Definition 4 and Algorithm 1] In Algorithm 1, the singular-value threshold is described only as 'a small numerical threshold.' Since the analysis invokes the degenerate fallback and the norm projection, it would help to specify how the threshold relates to r_M from Proposition 3, or at least note that the implementation threshold is not part of the theoretical guarantee.
  4. [Section VII, Table IV and Figure 6] The λ-sweep is described as 'flat' over four decades, but the gap changes from -0.70% to -0.55%, which is a small but visible trend. A sentence noting this is within seed noise would be useful.
  5. [Appendix A] The heuristic contraction route is clearly labeled as a sketch, but the condition L̃_f < 1 is stated without proof; since the main text relies on (B2) as a primitive, this is acceptable, but a pointer to where the rigorous route might appear would help readers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central guarantee is a conditional theorem whose unverified stability/excitation premises are stated as open conditions, not fitted outputs or self-citation load-bearing steps.

full rationale

The paper's derivation chain is not circular. The main sample-complexity result, Theorem 6, is explicitly conditional on Assumption Sets A, B, C, and D, and the constants in the guarantee are derived from those primitives rather than fitted to match the experiments. The most delicate premise is the TD-stability lower bound sigma_min(M_i^theta) >= mu_M(m) > 0 (Assumption 4), reduced in Lemma 3 to lambda_min(H_i^theta) >= h_M(m) > 0. The paper itself flags that verifying h_M(m) > 0 for a concrete continuous networked model class 'is left open' (Section V, after Lemma 3) and labels both (C) and (C') 'unverified structural premises' (Contributions). This is a genuine correctness/verification gap, but it is not circular: the theorem does not define mu_M in terms of the quantity it predicts, nor does it claim to have verified the excitation bound. The proof of Lemma 5 explicitly checks that the projection radius W* depends only on primitives and not on the target epsilon or on random quantities, precluding the classical fitted-input-called-prediction pattern. The experiments are also candidly delimited: the main LCQ comparison uses a structured quadratic critic and is described as an implementation-level demonstration, while the RFF-LSTD critic is exercised separately in the m-sweep with acknowledged confounds (Section VII-C). No load-bearing conclusion rests on a self-citation: the references to Qu et al., Lin et al., and Ren et al. provide context and prior frameworks, while the decay assumptions (B2) and (D3)(ii) are adopted as primitives in the present paper rather than imported by citation as proofs. The paper's own limitation statements, including the vacuous early-stopping caveat for Theorem 7(a) and the post-hoc-only nature of the Proposition 3 certificate, are honest conditionality, not hidden circularity. Therefore the appropriate finding is no significant circularity, with a score of 0.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The central sample-complexity theorem depends on an assumed exponential-decay structure (B2), an n-uniform smoothness/decay regularity (D3), and a TD-stability/excitation premise (Assumption 4 / Lemma 3) that the paper itself labels unverified. The excitation constant h_M is not derived or verified for any concrete class; the dimensional obstruction only provides upper bounds. These are the ledger entries a reader pays for upstream.

free parameters (2)
  • mu_M(m) = unknown; postulated > 0
    Uniform TD-stability lower bound in Assumption 4. All sample complexities and floor constants scale with 1/mu_M^2, but no concrete positive value or model class is provided; only an O(m^{-1/2}) upper bound is proved in Proposition 2.
  • h_M(m) = unknown; postulated > 0
    Feature-excitation lower bound in Lemma 3, implying mu_M >= (1 - sqrt(gamma)) h_M. The paper explicitly states verification for a concrete continuous networked model class is left open.
assumptions (6)
  • domain assumption (A1) Local Gaussian transitions with bounded drift and bounded initial support
    Assumption Set A, Section II-C. Supplies boundedness of the splitting factor and the Gaussian tail structure used throughout the RFF and confinement analysis.
  • domain assumption (A3) Positive-density bounded-score local policies, W2-Lipschitz in local state
    Assumption Set A, Section II-C. Justifies bounded-gradient concentration and the total-variation continuity argument for the population TD matrix in Lemma 10.
  • domain assumption (B2) Exponential decay of Q-functions: |Q_pi_i(s,a) - Q_pi_i(s',a')| <= c rho^{kappa+1} on kappa-neighborhood agreement
    Assumption Set B, Section II-C. The load-bearing locality hypothesis of the whole paper; Appendix A gives only a heuristic contraction sketch, explicitly not a proof.
  • domain assumption (D3)(ii)-(iii) Differentiated curvature decay and graph-tail (Delta-1)rho < 1
    Assumption Set D, Section II-C. Needed for the n-uniform bound on L_J in Appendix B; the paper states this does not follow from (B2).
  • ad hoc to paper Assumption 4 (Set C): uniform TD-stability sigma_min(M_theta_i) >= mu_M(m) > 0 with probability at least 1 - delta_M(m)
    Section V, Assumption 4. The paper labels it an “unverified structural premise”; the probability is postulated, not sample-controlled. Theorem 6 is stated under it.
  • ad hoc to paper Assumption 5 (Set C'): trajectory-conditional TD-stability on the realized sample path
    Section V, Assumption 5. Used only for the companion Theorem 7; the Proposition 3 certificate is post-hoc and does not instantiate the assumption prospectively.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Scalable Policy Optimization for Networked Multi-Agent Reinforcement Learning with Continuous State-Action Spaces." pith.science (2026). https://pith.science/paper/HFN4DNYS

@misc{pith2026260718554,
  author       = {Pith},
  title        = {Pith review of: Scalable Policy Optimization for Networked Multi-Agent Reinforcement Learning with Continuous State-Action Spaces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HFN4DNYS}},
  note         = {Machine review of arXiv:2607.18554}
}
abstract

We develop the Continuous Distributed Coupled Policy Gradient (CDCPG) algorithm for cooperative reinforcement learning in networked Markov decision processes with continuous state and action spaces. Each agent maintains a local actor over a bounded graph neighborhood, and a localized least-squares temporal-difference critic evaluates a truncated action-value function through a spectral random-feature representation of the local transition kernel. The analysis makes four contributions. First, the truncated action-value function is constructed as a conditional expectation over the neighborhood, yielding a well-posed localized Bellman theory that removes the continuation-kernel mismatch of naive truncation arguments. Second, we expose a dimensional obstruction to temporal-difference stability for normalized random features and prove an unconditional excitation bound that reduces stability to a symmetric persistence-of-excitation condition, monitorable through an online matrix-concentration certificate. Third, under exponential spatial decay of agent interactions, the excitation condition, and smoothness of the objective, CDCPG drives an averaged per-agent stationarity measure to within any excess $\epsilon$ of an explicitly characterized approximation floor using $\widetilde{\mathcal{O}}(\epsilon^{-2})$ shared-oracle samples, and the excess dependence matches the smooth nonconvex first-order rate; per-agent computation and communication are governed by the neighborhood size rather than the network size. Fourth, an adaptive-locality rule selects the radius that balances truncation and graph-decay residuals against the target accuracy. Experiments on a networked linear-quadratic benchmark corroborate the locality and feature-dimension predictions.

Figures

Figures reproduced from arXiv: 2607.18554 by the authors.

Figure 1
Figure 1. Per-agent discounted return J¯(θk) versus iteration k on LCQ (n = 9, 5 seeds, mean ± one std). Dashed reference V ref σ = −2.713. CDCPG matches the Centralized full-information baseline within 0.11 percentage points at 5 seeds. The comparison uses the structured quadratic-basis critic of Section VII-A; the analyzed RFF–LSTD critic is exercised in [PITH_FULL_IMAGE:figures/full_fig_p017_1.png] view at source ↗
Figure 3
Figure 3. Final-iteration per-agent return on LCQ, mean [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Relative Q-function approximation error [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Empirical absolute cross-agent reward covariance [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: CDCPG ablation on LCQ (K = 150, 3 seeds; mean ± one std). Dashed reference: V ref σ = −2.713. (a) Truncation￾radius sweep with consecutive-hop ratios ≈ 0.5, consistent with Lemma 2. (b) Regularization sweep with flat regime over four decades and break at λ = 1, consist…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 2 linked inside Pith

  1. [1]

    Multi-agent reinforcement learning: A selective overview of theories and algorithms,

    K. Zhang, Z. Yang, and T. Bas ¸ar, “Multi-agent reinforcement learning: A selective overview of theories and algorithms,”Handbook of rein- forcement learning and control, pp. 321–384, 2021

  2. [2]

    Stability constrained reinforcement learning for decentralized real-time voltage control,

    J. Feng, Y . Shi, G. Qu, S. H. Low, A. Anandkumar, and A. Wierman, “Stability constrained reinforcement learning for decentralized real-time voltage control,”IEEE Transactions on Control of Network Systems, vol. 11, no. 3, pp. 1370–1381, 2023

  3. [3]

    Scalable reinforcement learning of localized policies for multi-agent networked systems,

    G. Qu, A. Wierman, and N. Li, “Scalable reinforcement learning of localized policies for multi-agent networked systems,” inLearning for Dynamics and Control. PMLR, 2020, pp. 256–266

  4. [4]

    Scalable reinforcement learning for multiagent networked sys- tems,

    ——, “Scalable reinforcement learning for multiagent networked sys- tems,”Operations Research, vol. 70, no. 6, pp. 3601–3628, 2022

  5. [5]

    Multi-agent reinforcement learning in stochastic networked systems,

    Y . Lin, G. Qu, L. Huang, and A. Wierman, “Multi-agent reinforcement learning in stochastic networked systems,”Advances in neural informa- tion processing systems, vol. 34, pp. 7825–7837, 2021

  6. [6]

    Global convergence of localized policy iteration in networked multi-agent reinforcement learning,

    Y . Zhang, G. Qu, P. Xu, Y . Lin, Z. Chen, and A. Wierman, “Global convergence of localized policy iteration in networked multi-agent reinforcement learning,”Proceedings of the ACM on Measurement and Analysis of Computing Systems, vol. 7, no. 1, pp. 1–51, 2023

  7. [7]

    Fully decentralized multi-agent reinforcement learning with networked agents,

    K. Zhang, Z. Yang, H. Liu, T. Zhang, and T. Basar, “Fully decentralized multi-agent reinforcement learning with networked agents,” inInterna- tional conference on machine learning. PMLR, 2018, pp. 5872–5881

  8. [8]

    Finite-time analysis of dis- tributed td (0) with linear function approximation on multi-agent rein- forcement learning,

    T. Doan, S. Maguluri, and J. Romberg, “Finite-time analysis of dis- tributed td (0) with linear function approximation on multi-agent rein- forcement learning,” inInternational Conference on Machine Learning. PMLR, 2019, pp. 1626–1635

Show all 34 references
  1. [9]

    Decentralized online convex optimization in networked systems,

    Y . Lin, J. Gan, G. Qu, Y . Kanoria, and A. Wierman, “Decentralized online convex optimization in networked systems,” inInternational Conference on Machine Learning. PMLR, 2022, pp. 13 356–13 393

  2. [10]

    Multi-agent reinforcement learning for networked system control,

    T. Chu, S. Chinchali, and S. Katti, “Multi-agent reinforcement learning for networked system control,”arXiv preprint arXiv:2004.01339, 2020

  3. [11]

    Random features for large-scale kernel machines,

    A. Rahimi and B. Recht, “Random features for large-scale kernel machines,”Advances in neural information processing systems, vol. 20, 2007

  4. [12]

    Random features for ker- nel approximation: A survey on algorithms, theory, and beyond,

    F. Liu, X. Huang, Y . Chen, and J. A. Suykens, “Random features for ker- nel approximation: A survey on algorithms, theory, and beyond,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 10, pp. 7128–7148, 2021

  5. [13]

    Scalable spectral representations for multi-agent reinforcement learning in network mdps,

    Z. Ren, R. Zhang, B. Dai, and N. Li, “Scalable spectral representations for multi-agent reinforcement learning in network mdps,”arXiv preprint arXiv:2410.17221, 2024

  6. [14]

    Linear least-squares algorithms for temporal difference learning,

    S. J. Bradtke and A. G. Barto, “Linear least-squares algorithms for temporal difference learning,”Machine learning, vol. 22, no. 1, pp. 33– 57, 1996

  7. [15]

    Technical update: Least-squares temporal difference learning,

    J. A. Boyan, “Technical update: Least-squares temporal difference learning,”Machine learning, vol. 49, no. 2, pp. 233–246, 2002

  8. [16]

    Finite-sample analysis of least-squares policy iteration,

    A. Lazaric, M. Ghavamzadeh, and R. Munos, “Finite-sample analysis of least-squares policy iteration,”The Journal of Machine Learning Research, vol. 13, no. 1, pp. 3041–3074, 2012

  9. [17]

    A finite time analysis of temporal difference learning with linear function approximation,

    J. Bhandari, D. Russo, and R. Singal, “A finite time analysis of temporal difference learning with linear function approximation,” inConference on Learning Theory (COLT). PMLR, 2018, pp. 1691–1692

  10. [18]

    An introduction to matrix concentration inequalities,

    J. A. Tropp, “An introduction to matrix concentration inequalities,” Foundations and trends® in machine learning, vol. 8, no. 1-2, pp. 1–230, 2015

  11. [19]

    Vershynin,High-dimensional probability: An introduction with ap- plications in data science

    R. Vershynin,High-dimensional probability: An introduction with ap- plications in data science. Cambridge university press, 2018, vol. 47

  12. [20]

    Optimum bounds for the distributions of martingales in banach spaces,

    I. Pinelis, “Optimum bounds for the distributions of martingales in banach spaces,”The Annals of Probability, pp. 1679–1706, 1994

  13. [21]

    K. S. Narendra and A. M. Annaswamy,Stable adaptive systems. Courier Corporation, 2012

  14. [22]

    Policy gradi- ent methods for reinforcement learning with function approximation,

    R. S. Sutton, D. McAllester, S. Singh, and Y . Mansour, “Policy gradi- ent methods for reinforcement learning with function approximation,” Advances in neural information processing systems, vol. 12, 1999

  15. [23]

    Lower bounds for non-convex stochastic optimization,

    Y . Arjevani, Y . Carmon, J. C. Duchi, D. J. Foster, N. Srebro, and B. Woodworth, “Lower bounds for non-convex stochastic optimization,” Mathematical Programming, vol. 199, no. 1, pp. 165–214, 2023

  16. [24]

    On the theory of policy gradient methods: Optimality, approximation, and distribution shift,

    A. Agarwal, S. M. Kakade, J. D. Lee, and G. Mahajan, “On the theory of policy gradient methods: Optimality, approximation, and distribution shift,”Journal of Machine Learning Research, vol. 22, no. 98, pp. 1–76, 2021

  17. [25]

    Multi-agent actor-critic for mixed cooperative-competitive environ- ments,

    R. Lowe, Y . I. Wu, A. Tamar, J. Harb, O. Pieter Abbeel, and I. Mordatch, “Multi-agent actor-critic for mixed cooperative-competitive environ- ments,”Advances in neural information processing systems, vol. 30, 2017

  18. [26]

    Counterfactual multi-agent policy gradients,

    J. Foerster, G. Farquhar, T. Afouras, N. Nardelli, and S. Whiteson, “Counterfactual multi-agent policy gradients,” inProceedings of the AAAI conference on artificial intelligence, vol. 32, no. 1, 2018

  19. [27]

    Actor-attention-critic for multi-agent reinforcement learning,

    S. Iqbal and F. Sha, “Actor-attention-critic for multi-agent reinforcement learning,” inInternational conference on machine learning. PMLR, 2019, pp. 2961–2970

  20. [28]

    The surprising effectiveness of ppo in cooperative multi-agent games,

    C. Yu, A. Velu, E. Vinitsky, J. Gao, Y . Wang, A. Bayen, and Y . Wu, “The surprising effectiveness of ppo in cooperative multi-agent games,” Advances in neural information processing systems, vol. 35, pp. 24 611– 24 624, 2022

  21. [29]

    Near-optimal distributed linear-quadratic regulator for networked systems,

    S. Shin, Y . Lin, G. Qu, A. Wierman, and M. Anitescu, “Near-optimal distributed linear-quadratic regulator for networked systems,”SIAM Journal on Control and Optimization, vol. 61, no. 3, pp. 1113–1135, 2023

  22. [30]

    Network reconfiguration in distribution systems for loss reduction and load balancing,

    M. E. Baran and F. F. Wu, “Network reconfiguration in distribution systems for loss reduction and load balancing,”IEEE Transactions on Power delivery, vol. 4, no. 2, pp. 1401–1407, 1989

  23. [31]

    The description of a random field by means of conditional probabilities and conditions of its regularity,

    P. Dobruschin, “The description of a random field by means of conditional probabilities and conditions of its regularity,”Theory of Probability & Its Applications, vol. 13, no. 2, pp. 197–224, 1968

  24. [32]

    Can local particle filters beat the curse of dimensionality?

    P. Rebeschini and R. van Handel, “Can local particle filters beat the curse of dimensionality?”The Annals of Applied Probability, vol. 25, no. 5, pp. 2809–2866, 2015

  25. [33]

    Mini-batch stochastic approx- imation methods for nonconvex stochastic composite optimization,

    S. Ghadimi, G. Lan, and H. Zhang, “Mini-batch stochastic approx- imation methods for nonconvex stochastic composite optimization,” Mathematical Programming, vol. 155, no. 1, pp. 267–305, 2016

  26. [34]

    Stochastic first-and zeroth-order methods for nonconvex stochastic programming,

    S. Ghadimi and G. Lan, “Stochastic first-and zeroth-order methods for nonconvex stochastic programming,”SIAM journal on optimization, vol. 23, no. 4, pp. 2341–2368, 2013

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.