Pith. sign in

REVIEW 3 major objections 4 minor 29 references

Closed-Loop Bayesian Bandit Encoder with GRAND Receiver for a Bursty Interference Channel

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A receiver that learns the interference pattern makes non-interleaved transmission beat interleaving, with 5–10x lower block error rate in the evaluated configuration.

desk verdict A genuinely new closed-loop GRAND/bandit architecture with a plausible but under-tested preference reversal that rests on an asymmetric decoder comparison. read the letter →

arxiv 2607.15404 v2 pith:XWU6QESS submitted 2026-07-16 cs.IT cs.LGeess.SPmath.IT

Bhaskar Krishnamachari This is my paper
classification cs.ITcs.LGeess.SPmath.IT
keywords GRANDinterleavingburstyinterferencehiddenMarkovmodelThompsonsamplingclosed-loopadaptationblockerrorraterandomlinearcode
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 asks whether interleaving is always the right response to bursty interference, given a decoder that can learn the channel's temporal structure. It builds a closed loop in which the transmitter chooses per packet between an interleaved and a non-interleaved version of the same random linear code, while the receiver decodes with GRAND and a replaceable noise model. The transmitter estimates the number, amplitudes, and timing of unknown on/off interferers from aggregate receiver feedback, and once confident, the receiver switches to a hidden-Markov-based noise model that orders GRAND queries by posterior bit-flip probabilities. The paper's central claim is that the preferred arm reverses mid-connection: interleaving wins before the receiver adapts, but after the learned decoder is activated the non-interleaved arm achieves 5–10x lower block error rate. This matters because it shows that receiver-side adaptation can reverse the reliability–latency tradeoff that interleaving is usually meant to solve.

What carries the argument

The central object is GRAND (Guessing Random Additive Noise Decoding), whose codebook test and noise model are modular: the code determines only which candidate noise patterns are valid, while the noise model determines the order in which they are tested. The learned noise model builds a hidden Markov model over interferer on/off states from the estimated parameters, computes per-bit posterior flip probabilities via forward–backward smoothing, and orders GRAND queries with an SGRAND-style max-heap under a factorized posterior approximation. A discounted Thompson sampler at the transmitter selects between the two arms using a goodput-minus-latency reward, and a model-informed variant injects

What would settle it

Run the closed-loop system with the interleaved arm also equipped with a learned noise model that accounts for the interleaver permutation; if the interleaved arm's post-adaptation block error rate then matches or beats the non-interleaved arm's, the claimed reversal is an artifact of asymmetric adaptation.

Watch

Extended reading notes

Core claim

In the evaluated closed-loop configuration, the preferred transmission mode changes as a result of the receiver's own learning. Before channel estimation converges, the interleaved arm earns higher reward; after the receiver switches to the learned HMM-ordered GRAND decoder, the non-interleaved arm becomes preferable, with block error rates 5–10x lower than the interleaved arm across five simulation seeds. The reversal is endogenous: it is the decoder adaptation, not a change in the channel, that alters the reward distribution of the non-interleaved arm. In the reference configuration the learned noise model reduces block error rate by approximately one order of magnitude relative to ORBGRAN

Load-bearing premise

Only the non-interleaved arm receives the learned HMM decoder; the interleaved arm continues to use ORBGRAND, so the central preference reversal is measured against an unadapted baseline rather than two equally adapted options.

Editorial extensions

If this is right

  • If the receiver can learn the interference structure, interleaving's reliability advantage disappears while its delay cost remains, so non-interleaved transmission becomes the preferred mode.
  • Adapting the GRAND noise model to the learned channel reduces block error rate by roughly an order of magnitude compared with an untuned ORBGRAND decoder on the same channel.
  • Using partial channel estimates before full convergence reduces pre-convergence block error rate by up to 4.5x, so early adaptation is beneficial even before the estimator is confident.
  • Confidence-weighted model predictions accelerate the preference reversal by 23–38% and reduce post-transition selection of the lower-reward arm by about 65%.
  • Under a 100 MHz 5G NR-like symbol rate, the learning transient corresponds to a few milliseconds of occupied symbol time, suggesting the mechanism can operate within existing control-loop timescales.

Reading between the lines

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

  • The measured preference reversal compares an adapted non-interleaved arm against a fixed ORBGRAND interleaved arm; if the interleaved arm also received a learned, interleaver-aware noise model, the post-adaptation gap could shrink or vanish. The paper explicitly leaves this out of scope, but the headline claim depends on this asymmetry.
  • The reward's latency weight c_l is a free design parameter; the boundary between the two arms likely shifts with c_l, so the claimed reversal is a statement about this specific configuration, not a universal criterion for disabling interleaving.
  • The same closed loop could be applied to a richer family of transmission modes (different interleaver depths, code rates, or MCS entries), with the bandit selecting among them while the receiver adapts its noise model for each; whether the reversal persists across code parameters is a testable extension.
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 / 4 minor

Summary. The paper proposes a closed-loop system for packet-level selection between a non-interleaved random linear code and the same code with cross-codeword interleaving over an AWGN channel with unknown on/off interferers. The receiver uses GRAND with a replaceable noise model; aggregate channel statistics are fed back to a Bayesian learner at the transmitter, which estimates interference amplitudes/duty cycles and, once confident, enables an HMM-based noise model that computes posterior per-bit flip probabilities and orders GRAND queries. A discounted Thompson sampler chooses the transmission mode using a goodput-minus-latency reward. Simulations in five phases show that the interleaved arm is preferred before the receiver adapts and the non-interleaved arm becomes preferred after the learned HMM decoder is activated, with 5–10× lower BLER across five seeds; model-informed pseudo-observations accelerate the switch and reduce post-transition selection of the inferior arm.

Significance. If the central claim holds, the paper makes a useful conceptual contribution: receiver-side decoder adaptation can alter the relative value of interleaved versus non-interleaved transmission, so interleaving policy, channel learning, and decoder design should be treated as coupled. The paper is careful in its structure: V0 validates GRAND against an analytic abandonment bound, V1 provides oracle-parameter decoder comparisons, and the limitation statements in Sections V-E, VI-C, and XI are unusually explicit. The five-seed consistency is suggestive. However, the headline preference reversal is evaluated against an asymmetric decoder baseline, and the quantitative support is thin—five seeds, one operating point, point estimates, and no code/data—so the result is plausible but not yet established at the strength claimed in the abstract.

major comments (3)
  1. [Section VI-D and Table VI] The post-adaptation comparison is asymmetric. Section VI-D states that 'only arm A receives the learned model. Arm B continues to use ORBGRAND.' Table VI's 5–10× BLER advantage of arm A over arm B therefore bundles the arm choice with the decoder choice: the non-interleaved arm gets the learned HMM ordering, while the interleaved arm is stuck with a fixed generic decoder. The paper asserts the interleaver weakens temporal dependence but does not quantify this. Since the receiver already collects statistics in channel order for arm B (Section V-A), an interleaver-aware learned decoder for arm B is a natural matched baseline and is explicitly deferred as out of scope. Without that baseline, the preference reversal may be partly a decoder-capability artifact rather than a property of non-interleaved transmission. Please either add the matched baseline or reframe the claim as 'non-interleave
  2. [Section IX-E, Table VI] The statistical support for the headline claim needs strengthening. Table VI reports BLER over the full run, but arm B is played predominantly before the confidence switch and arm A predominantly after it, so the full-run 5–10× BLER ratio conflates the pre- and post-adaptation regimes; the per-arm separation is shown only for seed 0 in Fig. 5(b). The claim that non-interleaved transmission is preferable after decoder activation needs per-arm BLER measured over the post-switch interval for all five seeds, with error counts and Wilson intervals. Additionally, all numerical results rest on five seeds at a single Eb/N0, and the manuscript provides no code or data availability statement, which limits reproducibility of the simulation-based claims.
  3. [Eq. (7) and Section XI] The 'reliability–latency tradeoff reversal' depends on how latency enters the reward. The reward in Eq. (7) collapses interleaving latency to a constant penalty c_l D with c_l chosen so the penalty is about 10% of maximum goodput; delayed acknowledgments, retransmission timing, and scheduling delays are not modeled, as Section XI acknowledges. Since no sensitivity sweep over c_l is provided, the reward-based preference reversal is conditional on this specific design weight. The paper should either provide a c_l sensitivity analysis or state explicitly that the post-adaptation advantage is primarily a BLER effect, which would make the latency modeling less consequential for the central claim.
minor comments (4)
  1. [Abstract and Table VII] The abstract's 'approximately 65%' reduction in post-transition selection of the inferior arm is based on three seeds in V3b and the percentage is not derived from the tabulated shares. Please state the calculation or soften the claim.
  2. [Section X] The airtime conversion treats the BPSK stream as occupying 100 Msym/s after deriving roughly 92 Msym/s from the 5G NR numerology. Clarify that 100 Msym/s is an idealization used for the top axes.
  3. [References] Reference [3] contains stray spaces in the author initials ('V . Y . Wei'); please fix formatting. There are also a few inconsistent spaces in the bibliography that copyediting should catch.
  4. [Section V-B] The claim that EM failure 'suggests that the difficulty is statistical rather than solely an initialization failure' is reasonable but appears to be based on a small number of fits. A sentence describing the number of initializations and convergence criterion would be helpful.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the headline preference reversal is a simulation result driven by real ACK/NACK rewards, not by construction from the paper's assumptions.

full rationale

The paper's central claims are empirical simulation outcomes rather than derivations that reduce to their inputs. The post-adaptation preference for the non-interleaved arm is evaluated on actual simulated ACK/NACK rewards; the HMM noise model is built from estimated channel parameters (amplitudes, duty cycles, ON lengths) and the GRAND query ordering is a stated factorized approximation (Section VI-C), not a quantity fitted to the arm-preference outcome. Oracle comparisons use true channel parameters as independent benchmarks (Section IX-B). The model-informed Thompson sampler injects Monte Carlo predicted utilities computed from the same channel estimate (Section VII-A), but the reported benefit is measured against real rewards and is not a renamed fit: the pseudo-observations are confidence-weighted predictions, not fitted constants reproducing the claimed improvement. The only self-citation is reference [29] in a future-work enumeration of bandit algorithms; it is not load-bearing for any result. The paper explicitly acknowledges that only arm A receives the learned HMM decoder while arm B continues to use ORBGRAND (Section VI-D), and Section XI lists this and other limitations; this asymmetry is a threat to external validity or fairness of comparison, not circularity, because it does not make the observed preference equal to the model assumption by construction. No step was found where a prediction is definitionally identical to a fitted input or where the argument relies on an unverified self-citation chain.

Assumptions & free parameters 11 free parameters · 7 assumptions · 0 invented entities

The central claims rest on a hand-parameterized simulation of a synthetic channel; no code or data is shipped. There are no new physical entities. The load-bearing assumptions are the single-active-interferer HMM, the factorized ordering approximation, and the asymmetric decoder treatment of the two arms.

free parameters (11)
  • latency weight c_l = 0.005
    Sets arm-B interleaving delay penalty to ~10% of max goodput; arm ranking/reversal depends on it; sensitivity deferred to future work.
  • interleaver depth D = 16
    Smallest power of two >= max interferer ON length; defines arm-B latency and burst spreading.
  • run threshold tau = 1.5 * sigma_hat
    Impaired-run detection threshold; detection/false-run tradeoff.
  • min run length = 3
    Suppresses false runs from noise in run statistics.
  • query budget Q = 1e6 (V0), 1e5 (loop)
    GRANDAB abandonment bound; sets BLER floor through tail probability.
  • feedback period = 200 packets
    Cadence of aggregate statistics; sets estimation resolution and confidence updates.
  • TS discount gamma = 0.99
    Effective memory ~100 packets; controls reversal timing and exploration.
  • pseudo-count m0 = 30
    Strength of model-informed pseudo-observations in Normal-Gamma posterior.
  • confidence thresholds = S>100, c>=2, w>=0.15, Hausdorff <0.15
    Heuristic confidence gate (Eq. 4); hand-designed, not derived.
  • operating Eb/N0 = 6 dB
    Single reference SNR; no sweep across SNR reported.
  • code parameters = (128,105)
    High-rate random linear code from GRAND literature; fixes error floor ~2^-23.
assumptions (7)
  • standard math GRAND first-hit is ML if query order matches channel statistics
    From ref [1]; underlies all claims that learned ordering improves decoding.
  • domain assumption Interference model: k on/off interferers, exponential OFF, fixed ON length and amplitude, BPSK/AWGN
    Sec. IV-B; synthetic model, all results conditional on it.
  • domain assumption HMM state space assumes at most one active interferer per symbol
    Sec. VI-A; overlap probability ~4e-4 in studied regime; breaks at higher duty cycles.
  • ad hoc to paper GRAND ordering uses factorized posterior approximation, not full joint HMM posterior
    Sec. VI-C; exactness only w.r.t. factorized model.
  • domain assumption Arm B (interleaved) is not given a learned model; stays ORBGRAND
    Sec. VI-D; reversal claim depends on this asymmetry.
  • domain assumption Feedback link error-free, rate-limited, with no airtime cost
    Sec. IV/XI; idealized control plane.
  • domain assumption Statistics are collected only from successfully decoded packets
    Sec. IV/XI; acknowledged bias.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Closed-Loop Bayesian Bandit Encoder with GRAND Receiver for a Bursty Interference Channel." pith.science (2026). https://pith.science/paper/XWU6QESS

@misc{pith2026260715404,
  author       = {Pith},
  title        = {Pith review of: Closed-Loop Bayesian Bandit Encoder with GRAND Receiver for a Bursty Interference Channel},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XWU6QESS}},
  note         = {Machine review of arXiv:2607.15404}
}
abstract

Interleaving mitigates burst errors but introduces decoding delay and removes temporal error structure that a channel-aware decoder could exploit. We consider packet-level selection between a random linear code and the same code used with cross-codeword interleaving, over a channel with an unknown number of on/off interferers. The receiver uses Guessing Random Additive Noise Decoding (GRAND) with a replaceable noise model and feeds aggregate channel statistics back to a Bayesian estimator at the transmitter. Once the interference amplitudes and timing parameters are estimated, the receiver's noise model is replaced: it computes hidden-Markov-model posterior bit-flip probabilities and uses them to order GRAND queries. A discounted Thompson sampler selects between the two transmission modes using a goodput-minus-latency reward whose distribution is endogenously nonstationary: receiver adaptation, rather than channel change, alters the value of each mode. Across five simulation seeds, the interleaved mode is preferred before channel estimation converges. After the learned decoder is activated, the non-interleaved mode becomes preferable because it achieves lower block error rate without interleaving delay. In the reference configuration, the learned noise model reduces block error rate by approximately one order of magnitude relative to ORBGRAND. Using partial channel estimates before full convergence reduces pre-convergence block error rate by up to $4.5\times$. Adding model-predicted utilities as confidence-weighted pseudo-observations reduces post-transition selection of the inferior arm by approximately $65\%$. Under an idealized airtime conversion at a 100~MHz 5G~NR-like symbol rate, the learning transient corresponds to a few milliseconds of occupied symbol time.

Figures

Figures reproduced from arXiv: 2607.15404 by the authors.

Figure 1
Figure 1. System block diagram. Arm A is the non-interleaved RLC [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. overlays the simulated block error rate of hard￾detection GRANDAB on AWGN against two references: the finite-blocklength normal approximation for the induced binary symmetric channel, and the abandonment bound P(#errors > wcov), where wcov = 3 is the deepest error weight fully covered by the 106 -query budget. The simula￾tion follows the abandonment bound across four decades (at 6.5 dB, 1.11 × 10−3 simulated, 100 er… view at source ↗
Figure 3
Figure 3. V2: closed-loop estimation with arm A fixed (12,000 packets, [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: V2b: BLER under the three gating policies for the reference seed, [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: V3 results for seed 0 over 20,000 packet decisions. (a) Estimated [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 4 linked inside Pith

  1. [1]

    Capacity-achieving guessing random additive noise decoding,

    K. R. Duffy, J. Li, and M. M ´edard, “Capacity-achieving guessing random additive noise decoding,”IEEE Transactions on Information Theory, vol. 65, no. 7, pp. 4023–4040, 2019

    2019

  2. [2]

    Ordered reliability bits guessing random additive noise decoding,

    K. R. Duffy, W. An, and M. M ´edard, “Ordered reliability bits guessing random additive noise decoding,”IEEE Transactions on Signal Process- ing, vol. 70, pp. 4528–4542, 2022

    2022

  3. [3]

    ORBGRAND is almost capacity-achieving,

    M. Liu, Y . Wei, Z. Chen, and W. Zhang, “ORBGRAND is almost capacity-achieving,”IEEE Transactions on Information Theory, vol. 69, no. 5, pp. 2830–2840, 2023

    2023

  4. [4]

    High-throughput and energy-efficient VLSI architecture for ordered reliability bits GRAND,

    S. M. Abbas, T. Tonnellier, F. Ercan, M. Jalaleddine, and W. J. Gross, “High-throughput and energy-efficient VLSI architecture for ordered reliability bits GRAND,”IEEE Transactions on Very Large Scale Integration (VLSI) Systems, vol. 30, no. 6, pp. 681–693, 2022

    2022

  5. [5]

    A sub-0.8-pj/bit 16.3-gbps/mm 2 universal soft-detection decoder using ORBGRAND in 40-nm CMOS,

    A. Riaz, A. Yasar, F. Ercan, W. An, J. Ngo, K. Galligan, M. M ´edard, K. R. Duffy, and R. T. Yazicigil, “A sub-0.8-pj/bit 16.3-gbps/mm 2 universal soft-detection decoder using ORBGRAND in 40-nm CMOS,” inProceedings of the IEEE International Solid-State Circuits Conference (ISSCC), 2023

    2023

  6. [6]

    Soft maximum likelihood decoding using GRAND,

    A. Solomon, K. R. Duffy, and M. M ´edard, “Soft maximum likelihood decoding using GRAND,” inProceedings of the IEEE International Conference on Communications (ICC), 2020

    2020

  7. [7]

    Keep the bursts and ditch the interleavers,

    W. An, M. M ´edard, and K. R. Duffy, “Keep the bursts and ditch the interleavers,”IEEE Transactions on Communications, vol. 70, no. 6, pp. 3655–3667, 2022

    2022

  8. [8]

    Guessing random additive noise decoding with symbol reliability information (SRGRAND),

    K. R. Duffy, M. M ´edard, and W. An, “Guessing random additive noise decoding with symbol reliability information (SRGRAND),”IEEE Transactions on Communications, vol. 70, no. 1, pp. 3–18, 2022

    2022

Show all 29 references
  1. [9]

    GRAND for Gaussian intersymbol interference channels,

    Z. Li and W. Zhang, “GRAND for Gaussian intersymbol interference channels,”arXiv preprint arXiv:2603.08325, 2026

    2026

  2. [10]

    Decoding in the presence of ISI without interleaving – ORBGRAND-AI,

    K. R. Duffy, M. Grundei, J. A. Millward, M. Rangaswamy, and M. M ´edard, “Decoding in the presence of ISI without interleaving – ORBGRAND-AI,”arXiv preprint arXiv:2510.14939, 2025

    2025 arXiv

  3. [11]

    Universal decoding over finite-state additive channels via noise guessing,

    H. K. Miyamoto and S. Yang, “Universal decoding over finite-state additive channels via noise guessing,”arXiv preprint arXiv:2501.12971, 2025

    2025 arXiv

  4. [12]

    Joint error correction and fading channel estimation enhancement leveraging GRAND,

    C. Wiame, K. R. Duffy, and M. M ´edard, “Joint error correction and fading channel estimation enhancement leveraging GRAND,”arXiv preprint arXiv:2506.14756, 2025

    2025 arXiv

  5. [13]

    Capacity of a burst-noise channel,

    E. N. Gilbert, “Capacity of a burst-noise channel,”Bell System Technical Journal, vol. 39, no. 5, pp. 1253–1265, 1960

    1960

  6. [14]

    Estimates of error rates for codes on burst-noise channels,

    E. O. Elliott, “Estimates of error rates for codes on burst-noise channels,” Bell System Technical Journal, vol. 42, no. 5, pp. 1977–1997, 1963

    1977

  7. [15]

    Models for channels with memory and their applications to error control,

    L. N. Kanal and A. R. K. Sastry, “Models for channels with memory and their applications to error control,”Proceedings of the IEEE, vol. 66, no. 7, pp. 724–744, 1978

    1978

  8. [16]

    Cognitive radio: Brain-empowered wireless communica- tions,

    S. Haykin, “Cognitive radio: Brain-empowered wireless communica- tions,”IEEE Journal on Selected Areas in Communications, vol. 23, no. 2, pp. 201–220, 2005

    2005

  9. [17]

    Cognitive medium access: Exploration, exploitation, and competition,

    L. Lai, H. E. Gamal, H. Jiang, and H. V . Poor, “Cognitive medium access: Exploration, exploitation, and competition,”IEEE Transactions on Mobile Computing, vol. 10, no. 2, pp. 239–253, 2011

    2011

  10. [18]

    Optimal rate sampling in 802.11 systems: Theory, design, and implementation,

    R. Combes, J. Ok, A. Prouti `ere, D. Yun, and Y . Yi, “Optimal rate sampling in 802.11 systems: Theory, design, and implementation,”IEEE Transactions on Mobile Computing, vol. 18, no. 5, pp. 1145–1158, 2019

    2019

  11. [19]

    Rate adaptation with Thompson sampling in 802.11ac WLAN,

    H. Qi, Z. Hu, X. Wen, and Z. Lu, “Rate adaptation with Thompson sampling in 802.11ac WLAN,”IEEE Communications Letters, vol. 23, no. 10, pp. 1888–1892, 2019

    2019

  12. [20]

    Bayesian link adaptation under a BLER target,

    V . Saxena and J. Jald´en, “Bayesian link adaptation under a BLER target,” inProceedings of the IEEE 21st International Workshop on Signal Processing Advances in Wireless Communications (SPAWC), 2020

    2020

  13. [21]

    On the likelihood that one unknown probability exceeds another in view of the evidence of two samples,

    W. R. Thompson, “On the likelihood that one unknown probability exceeds another in view of the evidence of two samples,”Biometrika, vol. 25, no. 3–4, pp. 285–294, 1933

    1933

  14. [22]

    A tutorial on Thompson sampling,

    D. Russo, B. V . Roy, A. Kazerouni, I. Osband, and Z. Wen, “A tutorial on Thompson sampling,”Foundations and Trends in Machine Learning, vol. 11, no. 1, pp. 1–96, 2018

    2018

  15. [23]

    On upper-confidence bound policies for switching bandit problems,

    A. Garivier and E. Moulines, “On upper-confidence bound policies for switching bandit problems,” inProceedings of the International Conference on Algorithmic Learning Theory (ALT), 2011

    2011

  16. [24]

    Link adaptation using Joint-Thompson sampling,

    V . Vinjam, M. Kolavennu, M. Vajha, and K. P. Narayanaprasad, “Link adaptation using Joint-Thompson sampling,”arXiv preprint arXiv:2607.11075, 2026

    2026 arXiv

  17. [25]

    Soft-output (SO) GRAND and iterative decoding to outperform LDPC codes,

    P. Yuan, M. M ´edard, K. Galligan, and K. R. Duffy, “Soft-output (SO) GRAND and iterative decoding to outperform LDPC codes,”IEEE Transactions on Wireless Communications, 2025

    2025

  18. [26]

    Finite-time analysis of the multiarmed bandit problem,

    P. Auer, N. Cesa-Bianchi, and P. Fischer, “Finite-time analysis of the multiarmed bandit problem,”Machine Learning, vol. 47, no. 2–3, pp. 235–256, 2002

    2002

  19. [27]

    Deterministic sequencing of exploration and exploitation for multi-armed bandit problems,

    S. Vakili, K. Liu, and Q. Zhao, “Deterministic sequencing of exploration and exploitation for multi-armed bandit problems,”IEEE Journal of Selected Topics in Signal Processing, vol. 7, no. 5, pp. 759–767, 2013

    2013

  20. [28]

    The non- stochastic multiarmed bandit problem,

    P. Auer, N. Cesa-Bianchi, Y . Freund, and R. E. Schapire, “The non- stochastic multiarmed bandit problem,”SIAM Journal on Computing, vol. 32, no. 1, pp. 48–77, 2002

    2002

  21. [29]

    Combinatorial network op- timization with unknown variables: Multi-armed bandits with linear rewards and individual observations,

    Y . Gai, B. Krishnamachari, and R. Jain, “Combinatorial network op- timization with unknown variables: Multi-armed bandits with linear rewards and individual observations,”IEEE/ACM Transactions on Net- working, vol. 20, no. 5, pp. 1466–1478, 2012

    2012

Pith tools

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