pith. sign in

arxiv: 2606.06917 · v1 · pith:TJH4QVHEnew · submitted 2026-06-05 · 💻 cs.ET

Belief-Aware Scheduling for Predictive Wildfire Hazard Mapping under Sparse-Window Telemetry

Xun Shao , Kohsuke Yamakawa , Cheah Wai Shiang This is my paper

Pith reviewed 2026-06-27 20:22 UTC · model grok-4.3

classification 💻 cs.ET
keywords wildfire hazard mappingsparse-window telemetrybelief-aware schedulingpredictive mappingcross-region attention encoderpartially observed allocationsynthetic environment evaluationedge node monitoring
0
0 comments X

The pith

A structured belief derived from the H-step prediction requirements and maintained by a non-myopic scheduler lets a 40k-parameter cross-region attention encoder exceed activity-paced baselines by 28% on default landscapes and 11% on structu

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

The paper argues that the key design issue for edge wildfire monitoring is not the neural architecture but how to build and sustain a structured belief that meets the receiver's future prediction needs under duty-limited or windowed downlinks. It formalizes the problem as a partially observed sequential allocation task whose three per-region action axes are sensing, representation, and transmission, with each belief component taken directly from the input requirements of the H-step forward operator. Evaluation occurs in a physics-calibrated synthetic environment that independently varies window period, capacity, horizon, and fuel composition. Three results follow: the performance gap between non-myopic and uniform pacing peaks at intermediate sparsity, ablating the structured belief switches the dominant error source between temporal staleness and static-risk prior, and the lightweight encoder outperforms the reference while a deeper Transformer does not.

Core claim

The operative design problem is not which neural architecture to use but how to derive a structured belief sufficient for the receiver's prediction task and maintain it through a scheduler that anticipates future transmission opportunities. Identifying these mechanisms requires independent control over window period P, per-window capacity C, predictive horizon H, and fuel composition, which is not separable in real-landscape data; evaluation on a physics-calibrated synthetic environment shows that a 40 k-parameter lightweight cross-region attention encoder exceeds the FAIR activity-paced reference by ~28% on the default landscape and ~11% on the structured landscape, while a deeper Transform

What carries the argument

The structured belief in a partially observed sequential allocation problem with three coupled per-region action axes (sensing, representation, transmission), each component derived from the H-step forward operator's input requirements.

If this is right

  • The gap between non-myopic activity-paced scheduling and uniform pacing is unimodal in window-period sparsity and peaks at intermediate spacing.
  • Ablating the structured belief flips the dominant operative component between temporal staleness on the default landscape and static-risk prior on the structured landscape.
  • The per-cell intensity belief component is redundant in both landscapes.
  • A deeper Transformer encoder yields no improvement in mean predictive loss over the 40k-parameter lightweight cross-region attention encoder and shows higher training-seed variance.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same belief-construction and scheduling approach could be tested on other sparse-telemetry prediction tasks such as flood or traffic monitoring where the forward operator inputs are known.
  • If the unimodal sparsity gap holds outside the synthetic setting, operators could deliberately choose intermediate window periods rather than pushing for either very frequent or very infrequent contacts.
  • The finding that a modest architectural bias suffices once the belief and scheduling are correctly posed suggests that future work can focus on refining the belief derivation rather than scaling model size.

Load-bearing premise

The physics-calibrated synthetic environment supplies a faithful test of the scheduling and belief mechanisms because real-landscape telemetry does not permit independent control of window period, capacity, horizon, and fuel composition.

What would settle it

Running the identical scheduler and lightweight encoder on real-landscape telemetry and checking whether the reported 28% and 11% gains over the activity-paced reference disappear once window period, capacity, horizon, and fuel composition are no longer independently controllable.

Figures

Figures reproduced from arXiv: 2606.06917 by Cheah Wai Shiang, Kohsuke Yamakawa, Xun Shao.

Figure 1
Figure 1. Figure 1: Scenario and scheduling principle (two panels, paper-specific differentiators only). [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Empirical room for the principle. The FAIR activity-paced reference–uniform gap is non-monotonic in window period P on the d* arm: 46 % at P = 1, peaking at 54 % at P = 2, 44 % at P = 4, and dropping to 10 % at P = 8. Both bars at each P use the same FAIR (belief-only) logic; only the per-window budget allocation differs. The relative gap is gap/Lu (= improvement of activity-paced over uniform pacing). Not… view at source ↗
Figure 3
Figure 3. Figure 3: Mechanism. Operative belief components flip with landscape structure. (a) Left, default landscape: removing the temporal-staleness components is the most costly (+0.023, dominant component high￾lighted). (b) Right, structured landscape (zoomed y-axis): the static-risk prior S becomes dominant (+0.004); the temporal-staleness components are uninformative or mildly harmful. The per-cell hazard estimate Ft re… view at source ↗
Figure 4
Figure 4. Figure 4: Realisability check across landscapes. Both panels use matched N = 5 training seeds for all four architectures; bars show mean ± SE across training seeds and yellow markers show per-seed values. (a) Default landscape: Lightw. Attn. (LH = 0.0670 ± 0.0031 SE) reaches the lowest mean and exceeds the FAIR activity-paced reference (Section V-A); Transformer (LH = 0.0735±0.0055 SE) is close in mean but has highe… view at source ↗
Figure 6
Figure 6. Figure 6: Budget-dependent realisation. At small budgets (280 ep) the sim￾pler MLP is competitive; at moderate budgets (1200 ep) the deeper Trans￾former is realised first; at sufficient budgets (5000 ep) the lightweight attention encoder realises the principle at modest cost. recovered from the under-trained regime); at the full budget the lightweight attention realises the principle at modest cost. The implication … view at source ↗
Figure 7
Figure 7. Figure 7: H-axis ablation at the 1200-ep budget, with MLP and Transformer as architectural endpoints. The Transformer’s gap closure remains above the FAIR activity-paced reference (green dashed) across all H ∈ {1, 3, 6}; the MLP plateaus and at H = 6 is effectively tied with this reference. Gap closure > 100 % is expected (see Section V-A). The lightweight encoder’s 5000-ep value at H = 3 is in Table V; see Table VI… view at source ↗
read the original abstract

An edge node monitoring a wildfire observes more than a duty-limited or windowed down-link can carry. The receiver must predict the H-step-ahead hazard map from whatever the link delivers. We argue the operative design problem is not which neural architecture to use but how to derive a structured belief sufficient for the receiver's prediction task and maintain it through a scheduler that anticipates future transmission opportunities. We formalize this as a partially observed sequential allocation problem with three coupled per-region action axes (sensing, representation, transmission), and derive each component of the structured belief from the H-step forward operator's input requirements. Identifying these mechanisms requires independent control over the window period P, per-window capacity C, predictive horizon H, and fuel composition, which is not separable in real-landscape data; we therefore evaluate on a physics-calibrated synthetic environment. Three empirical observations support the principle: the gap between a non-myopic activity-paced reference and uniform pacing is unimodal in window-period sparsity, peaking at intermediate spacing; ablating the structured belief, the dominant operative component flips between a default landscape (temporal staleness) and a structured landscape (static-risk prior), while the per-cell intensity belief is redundant in both; and a 40 k-parameter lightweight cross-region attention encoder exceeds the FAIR activity-paced reference by ~28% on the default landscape and ~11% on the structured landscape. A deeper Transformer encoder does not improve over the lightweight encoder in mean predictive loss and exhibits higher training-seed variance. Within this task class and regime, a modest architectural inductive bias suffices when the belief and the scheduling problem are correctly posed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper claims that for predictive wildfire hazard mapping from sparse-window telemetry, the operative problem is deriving a structured belief from the H-step forward operator and maintaining it via a non-myopic scheduler that anticipates transmission opportunities. It formalizes the setting as a partially observed sequential allocation problem with coupled sensing/representation/transmission actions per region, evaluates the resulting belief-aware scheduler on a physics-calibrated synthetic environment (chosen because real telemetry does not permit independent control of P, C, H and fuel composition), and reports three observations: the performance gap versus uniform pacing is unimodal in window-period sparsity; ablating the structured belief flips the dominant component between temporal staleness and static-risk prior; and a 40 k-parameter lightweight cross-region attention encoder outperforms the FAIR activity-paced reference by ~28 % on the default landscape and ~11 % on the structured landscape, while a deeper Transformer does not improve mean loss.

Significance. If the empirical observations hold under a validated testbed, the work would indicate that modest architectural inductive bias is sufficient once the belief state and scheduling objective are correctly posed, offering a design principle for resource-constrained edge nodes in disaster-monitoring applications. The explicit separation of belief components and the demonstration that per-cell intensity belief is redundant are potentially reusable insights for other sparse-telemetry prediction tasks.

major comments (2)
  1. [Abstract / evaluation design] Abstract and evaluation-design paragraph: the three central empirical claims (unimodal gap, belief-component flip, 28 %/11 % gains) rest exclusively on a physics-calibrated synthetic environment whose forward operator and noise model are never quantitatively matched to any real wildfire downlink trace (hazard-map statistics, transmission-error distributions, or event-rarity profiles). Because the manuscript itself states that real telemetry does not permit independent control of P, C, H and fuel composition, the absence of such matching makes it impossible to assess whether the reported dominance reversal and encoder superiority are artifacts of the testbed rather than evidence for the design principle.
  2. [Abstract] Abstract: concrete percentage gains and ablation outcomes are reported without error bars, without statistical significance tests, and without any description of how the synthetic calibration was validated against real fire-spread statistics. This undermines the reliability of the ~28 % and ~11 % margins and of the claim that the lightweight encoder “suffices.”

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major point below, clarifying the rationale for the synthetic testbed while agreeing to improve presentation and validation details where feasible.

read point-by-point responses

  1. Referee: [Abstract / evaluation design] Abstract and evaluation-design paragraph: the three central empirical claims (unimodal gap, belief-component flip, 28 %/11 % gains) rest exclusively on a physics-calibrated synthetic environment whose forward operator and noise model are never quantitatively matched to any real wildfire downlink trace (hazard-map statistics, transmission-error distributions, or event-rarity profiles). Because the manuscript itself states that real telemetry does not permit independent control of P, C, H and fuel composition, the absence of such matching makes it impossible to assess whether the reported dominance reversal and encoder superiority are artifacts of the testbed rather than evidence for the design principle.

    Authors: We appreciate the referee's concern. The manuscript explicitly motivates the synthetic environment because real telemetry precludes independent control of P, C, H and fuel composition—the very variables needed to isolate the effects of structured belief and non-myopic scheduling. The physics calibration follows established wildfire-spread models (e.g., Rothermel-based simulators) to ensure the forward operator and noise are representative, yet we acknowledge that a direct quantitative match to specific real downlink traces (hazard statistics, error distributions) is not supplied. This is a genuine limitation: obtaining matched real traces while retaining experimental control is infeasible. We will revise the evaluation-design paragraph to (i) restate the control requirement more prominently, (ii) add a dedicated calibration subsection describing the parameter fitting procedure and any qualitative alignment with published fire-spread statistics, and (iii) qualify the claims as demonstrating the design principle under controlled conditions rather than claiming immediate real-world transfer. revision: partial

  2. Referee: [Abstract] Abstract: concrete percentage gains and ablation outcomes are reported without error bars, without statistical significance tests, and without any description of how the synthetic calibration was validated against real fire-spread statistics. This undermines the reliability of the ~28 % and ~11 % margins and of the claim that the lightweight encoder “suffices.”

    Authors: We agree that the abstract presentation can be strengthened. The full manuscript reports results averaged over multiple random seeds and landscape realizations; we will (i) insert parenthetical error ranges or standard deviations for the reported percentages in the abstract, (ii) add a short clause noting that differences are statistically significant at p<0.05 under paired t-tests across seeds, and (iii) expand the calibration-validation description in the main evaluation section (with a brief cross-reference in the abstract). Space constraints prevent a full validation paragraph in the abstract itself, but these changes will improve transparency without altering the core claims. revision: yes

standing simulated objections not resolved
  • We cannot supply a quantitative match of the synthetic forward operator and noise model to real wildfire downlink traces, because real telemetry does not permit independent control of P, C, H and fuel composition.

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper formalizes the problem as a partially observed sequential allocation task and states that each component of the structured belief is derived from the H-step forward operator's input requirements; this modeling step is independent of the subsequent empirical measurements. All reported performance deltas (28%/11% gains, unimodal gap, belief-component flip) are obtained by comparing against an external activity-paced reference scheduler inside a synthetic generator whose parameters (P, C, H, fuel composition) are set independently of the encoder losses. No self-definitional equations, fitted inputs renamed as predictions, or load-bearing self-citations appear in the abstract or evaluation design. The synthetic testbed is presented as a necessary proxy precisely because real telemetry does not permit the required separation, but the comparison metrics remain externally anchored.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 1 invented entities

The central claim rests on the assumption that the synthetic generator faithfully reproduces the statistical structure that would govern real telemetry scheduling decisions; the structured belief itself is defined by reference to the forward operator's input requirements, which are not independently validated.

free parameters (2)
  • window period P
    Varied as an experimental factor in the synthetic environment to produce the unimodal gap observation.
  • per-window capacity C
    Controlled factor whose interaction with P produces the reported performance surface.
axioms (1)
  • domain assumption The physics-calibrated synthetic environment reproduces the essential statistical dependencies between telemetry sparsity and H-step hazard prediction error that would appear in real landscapes.
    Invoked to justify use of synthetic data when real data does not allow independent control of P, C, H, and fuel composition.
invented entities (1)
  • structured belief no independent evidence
    purpose: Sufficient statistic for the receiver's H-step prediction task, derived from the forward operator's input requirements
    Introduced as the central object that the scheduler must maintain; no independent falsifiable prediction outside the synthetic runs is supplied.

pith-pipeline@v0.9.1-grok · 5829 in / 1695 out tokens · 25785 ms · 2026-06-27T20:22:35.569603+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 15 canonical work pages · 3 internal anchors

  1. [1]

    Data acquisition filtering focused on optimizing transmission in a LoRaW AN network applied to the WSN forest monitoring system,

    T. Brito, B. F. Azevedo, J. a. Mendes, M. Zorawski, F. P. Fernandes, A. I. Pereira, J. Rufino, J. Lima, and P. Costa, “Data acquisition filtering focused on optimizing transmission in a LoRaW AN network applied to the WSN forest monitoring system,”Sensors, vol. 23, no. 3, p. 1282, 2023

    2023

  2. [2]

    Beyond transmitting bits: Context, seman- tics, and task-oriented communications,

    D. G ¨und¨uz, Z. Qin, I. E. Aguerri, H. S. Dhillon, Z. Yang, A. Yener, K. K. Wong, and C.-B. Chae, “Beyond transmitting bits: Context, seman- tics, and task-oriented communications,”IEEE J. Sel. Areas Commun., vol. 41, no. 1, pp. 5–41, 2023

    2023

  3. [3]

    Context-aware semantic communication for the wireless networks,

    G. Liu, Y . Liu, J. Wang, H. Du, D. Niyato, J. Kang, Z. Xiong, and A. Jamalipour, “Context-aware semantic communication for the wireless networks,”arXiv preprint arXiv:2505.23249, 2025

    work page arXiv 2025

  4. [4]

    Task-adaptive semantic communications with controllable diffusion-based data regeneration,

    F. Guo, A. Wijesinghe, S. Zhang, and Z. Ding, “Task-adaptive semantic communications with controllable diffusion-based data regeneration,” arXiv preprint arXiv:2505.07980, 2025

    work page arXiv 2025

  5. [5]

    Universal rate- distortion-classification representations for lossy compression,

    N. Nguyen, T. Nguyen, T. Nguyen, and B. Bose, “Universal rate- distortion-classification representations for lossy compression,” inProc. IEEE Inf. Theory Workshop (ITW), 2025, arXiv:2504.09025

    work page arXiv 2025

  6. [6]

    From reactive to active sensing: A survey on information gathering in decision-theoretic planning,

    T. S. Veiga and J. Renoux, “From reactive to active sensing: A survey on information gathering in decision-theoretic planning,”ACM Comput. Surv., vol. 55, no. 13s, 2023. SHAOET AL.: BELIEF-AW ARE SCHEDULING FOR PREDICTIVE WILDFIRE HAZARD MAPPING 12

    2023

  7. [7]

    POPGym: Benchmarking partially observable reinforcement learn- ing,

    S. Morad, R. Kortvelesy, M. Bettini, S. Liwicki, and A. Prorok, “POPGym: Benchmarking partially observable reinforcement learn- ing,” inProc. Int. Conf. Learning Representations (ICLR), 2023, arXiv:2303.01859

    work page arXiv 2023

  8. [8]

    MaxIn- foRL: Boosting exploration in reinforcement learning through informa- tion gain maximization,

    B. Sukhija, S. Coros, A. Krause, P. Abbeel, and C. Sferrazza, “MaxIn- foRL: Boosting exploration in reinforcement learning through informa- tion gain maximization,” inProc. Int. Conf. Learning Representations (ICLR), 2025, arXiv:2412.12098

    work page arXiv 2025

  9. [9]

    Real-world reinforcement learning of active percep- tion behaviors,

    E. S. Hu, J. Wang, X. Yuan, F. Luo, M. Li, G. Lambrechts, O. Rybkin, and D. Jayaraman, “Real-world reinforcement learning of active percep- tion behaviors,” inAdvances in Neural Information Processing Systems (NeurIPS), 2025, arXiv:2512.01188

    work page arXiv 2025

  10. [10]

    Decision Transformer: Reinforcement Learning via Sequence Modeling

    L. Chen, K. Lu, A. Rajeswaran, K. Lee, A. Grover, M. Laskin, P. Abbeel, A. Srinivas, and I. Mordatch, “Decision Transformer: Reinforcement learning via sequence modeling,” inAdvances in Neural Information Processing Systems (NeurIPS), 2021, arXiv:2106.01345

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  11. [11]

    Real-time detection of forest fires using FireNet- CNN and explainable AI techniques,

    G. M. I. Alam, N. Tasnia, T. Biswas, M. J. Hossen, S. A. Tanim, and M. S. U. Miah, “Real-time detection of forest fires using FireNet- CNN and explainable AI techniques,”IEEE Access, vol. 13, pp. 51 150– 51 181, 2025

    2025

  12. [12]

    Drone-based AI system for wildfire monitoring and risk prediction,

    C. A. S. Lelis, J. J. Roncal, L. Silveira, R. D. G. a. de Aquino, C. A. C. Marcondes, J. C. Marques, D. S. Loubach, F. A. N. Verri, V . V . Curtis, and D. G. De Souza, “Drone-based AI system for wildfire monitoring and risk prediction,”IEEE Access, vol. 12, pp. 139 865–139 882, 2024

    2024

  13. [13]

    Survey of AI-enabled IoT for wildfire prediction, detection, and post- operations,

    A. Badary, D. Chen, S. Wang, W. Li, S. Chen, C. Huang, and M. Aibin, “Survey of AI-enabled IoT for wildfire prediction, detection, and post- operations,”IEEE Access, vol. 14, pp. 9819–9837, 2026

    2026

  14. [14]

    ForestProtector: An IoT architecture integrating machine vision and deep reinforcement learning for efficient wildfire monitoring,

    K. Bonilla-Ormachea, H. Cuizaga, E. Salcedo, S. Castro, S. Fernandez- Testa, and M. Mamani, “ForestProtector: An IoT architecture integrating machine vision and deep reinforcement learning for efficient wildfire monitoring,” inProc. 11th Int. Conf. Automation, Robotics and Appli- cations (ICARA), 2025

    2025

  15. [15]

    Detecting wildfires on UA Vs with real-time segmentation trained by larger teacher models,

    J. Pesonen, T. Hakala, V . Karjalainen, N. Koivum ¨aki, L. Markelin, A.-M. Raita-Hakola, J. Suomalainen, I. P ¨ol¨onen, and E. Honkavaara, “Detecting wildfires on UA Vs with real-time segmentation trained by larger teacher models,” inProc. IEEE/CVF Winter Conf. Applications of Computer Vision (WACV), 2025, arXiv:2408.10843

    work page arXiv 2025

  16. [16]

    Detecting wildfire flame and smoke through edge computing using transfer learning enhanced deep learning models,

    G. Vazquez, S. Zhai, and M. Yang, “Detecting wildfire flame and smoke through edge computing using transfer learning enhanced deep learning models,”arXiv preprint arXiv:2501.08639, 2025

    work page arXiv 2025

  17. [17]

    A risk-aware UA V-edge service framework for wildfire monitoring and emergency response,

    Y . Huang, Z. Wang, and R. Buyya, “A risk-aware UA V-edge service framework for wildfire monitoring and emergency response,”arXiv preprint arXiv:2602.19742, 2026

    work page arXiv 2026

  18. [18]

    Position Paper: From Edge AI to Adaptive Edge AI

    F. Pittorino and M. Roveri, “Position paper: From edge AI to adaptive edge AI,”arXiv preprint arXiv:2604.07360, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  19. [19]

    WildfireSpreadTS: A dataset of multi-modal time series for wildfire spread prediction,

    S. Gerard, Y . Zhao, and J. Sullivan, “WildfireSpreadTS: A dataset of multi-modal time series for wildfire spread prediction,” inProc. Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2023

    2023

  20. [20]

    Meso- geos: A multi-purpose dataset for data-driven wildfire modeling in the Mediterranean,

    S. Kondylatos, I. Prapas, G. Camps-Valls, and I. Papoutsis, “Meso- geos: A multi-purpose dataset for data-driven wildfire modeling in the Mediterranean,”arXiv preprint arXiv:2306.05144, 2023

    work page arXiv 2023

  21. [21]

    Next day wildfire spread: A machine learning dataset to predict wildfire spreading from remote-sensing data,

    F. Huot, R. L. Hu, N. Goyal, T. Sankar, M. Ihme, and Y .-F. Chen, “Next day wildfire spread: A machine learning dataset to predict wildfire spreading from remote-sensing data,” inIEEE Transactions on Geoscience and Remote Sensing, 2022, arXiv:2112.02447

    work page arXiv 2022

  22. [22]

    Applications of deep reinforcement learning in communications and networking: A survey,

    N. C. Luong, D. T. Hoang, S. Gong, D. Niyato, P. Wang, Y .-C. Liang, and D. I. Kim, “Applications of deep reinforcement learning in communications and networking: A survey,”IEEE Commun. Surveys Tuts., vol. 21, no. 4, pp. 3133–3174, 2019

    2019

  23. [23]

    Online learning based efficient resource allocation for LoRaW AN network,

    R. Wang, W. Li, J. Ren, T. Song, X. Wang, S. Wang, and S. Xu, “Online learning based efficient resource allocation for LoRaW AN network,” arXiv preprint arXiv:2509.10493, 2025

    work page arXiv 2025

  24. [24]

    Policy Distillation

    A. A. Rusu, S. G. Colmenarejo, c. G ¨ulc ¸ehre, G. Desjardins, J. Kirk- patrick, R. Pascanu, V . Mnih, K. Kavukcuoglu, and R. Hadsell, “Policy distillation,” inProc. Int. Conf. Learning Representations (ICLR), 2016, arXiv:1511.06295

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  25. [25]

    Firefighter math—rate of spread,

    National Wildfire Coordinating Group, “Firefighter math—rate of spread,” 2023

    2023

  26. [26]

    RP002-1.0.4 LoRaW AN regional parameters,

    LoRa Alliance, “RP002-1.0.4 LoRaW AN regional parameters,” LoRa Alliance Technical Committee, Tech. Rep., 2022

    2022

  27. [27]

    AN1200.13: LoRa modem designer’s guide,

    Semtech Corporation, “AN1200.13: LoRa modem designer’s guide,” Semtech Corporation, Tech. Rep., 2013. TABLE IX DETERMINISTIC AND STOCHASTIC EVALUATION MEANS PER ARCHITECTURE PER LANDSCAPE,BOTH REPORTED AS MEAN±SE ACROSS TRAINING SEEDS(EACH PER-SEED ENTRY IS ITSELF A MEAN ACROSS40EVALUATION SEEDS). THE“BEST”COLUMN IS THE PER-SEED MINIMUM OF DET AND STOCH,T...

    2013