{"id":"d1f787a5-59ae-47c2-99d6-ad60995254c8","arxiv_id":"2501.03448","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A VoL/TLW-based PDQN scheduler for NOMA federated meta-learning outperforms DDPG, OMA, equal-weight, and random baselines in simulation.","lead":"A new task-oriented scheduling method for federated meta-learning over NOMA wireless networks ranks devices by a 'value of learning' score and uses deep reinforcement learning to choose who transmits and at what power and speed. The authors report higher accuracy and value of learning than four simpler baselines in a 10-device non-IID CIFAR-10 simulation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Problem (21) omits the per-device time and energy budget constraints that the system model states as mandatory, so the reported VoL gains may come from schedules that violate the stated maximum tolerable limits.","rationale":"The reader's weakest assumption focused on the realism of the simulation parameters and the calibration of VoL/TLW. My concern is different and more fundamental: the optimization problem (21) omits the hard per-device time and energy constraints that the system model explicitly requires. This is an internal inconsistency, not a question of parameter realism. If the constraints are missing, the objective can reward infeasible schedules, undermining the reported VoL comparisons. The reader's rationale did mention a related formulation gap (the one-shot MINLP vs trajectory-dependent accuracy), but not the missing budget constraints. I do not see an outright mathematical contradiction beyond this, and the empirical results could survive correction if the constraints were added and the comparisons were re-run. Hence the reader's CONDITIONAL verdict remains appropriate, but the specific justification should be updated to include this gap. My read does not change the verdict, so I mark it UNCHANGED.","tokens_in":9100,"tokens_out":8764,"duration_ms":87163,"concrete_test":"Re-run the simulations in Section IV and record, for every scheduled device in every round, whether T_cmp_n + T_com_n <= T_max_n and E_cmp_n + E_com_n <= E_max_n. If any proposed-scheme deployment violates these limits, add these two sets of constraints to problem (21), retrain the PDQN agent, and recompute Figs. 4 and 5. If the VoL gap over baselines shrinks or the reported FML accuracy advantage changes materially, the central claim is compromised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The system model in Section II-C states that \"the total time and energy consumed by each device during model training must not exceed its specified maximum limits.\" However, the optimization problem (21) contains only the box constraints (21b)-(21d) on z_n, p_n, and f_n; it does not include constraints such as T_cmp_n + T_com_n <= T_max_n or E_cmp_n + E_com_n <= E_max_n. The objective (21a) penalizes V_T^n = T/T_max_n and V_E^n = E_n/E_max_n, but these are soft penalty terms, not hard constraints. The PDQN agent can therefore choose actions that exceed the per-device budgets while still obtaining a positive VoL contribution, especially when the accuracy term eta1*V_A^n is large. Because the baselines (e.g., OMA) may effectively consume more time or energy, a fair comparison requires that all schemes respect the same hard budgets. As written, the reported total VoL advantage could be an artifact of allowing infeasible resource consumption for the proposed NOMA-based scheme. This directly threatens the central claim that the framework \"significantly outperforms\" baselines in total VoL.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a task-oriented federated meta-learning (FML) framework over NOMA networks. It introduces two metrics, Value of Learning (VoL) and Task-Level Weight (TLW), to capture per-device training value and task importance, and formulates a mixed-integer nonlinear program (Problem 21) that jointly optimizes device scheduling, transmit power, and computation frequency. A parameterized deep Q-network (PDQN) is developed to handle the hybrid discrete-continuous action space. Simulations on non-IID CIFAR-10 report that the proposed scheme outperforms OMA, equal-weight, DDPG, and random-resource-allocation baselines in both FML accuracy and total VoL.","tokens_in":9499,"tokens_out":5285,"duration_ms":50922,"significance":"If the claims are validated, the framework is a useful task-aware design for wireless federated meta-learning, and the VoL/TLW metrics plus the PDQN solution are a coherent way to handle diverse device requirements. The paper gives a clear system model and MDP formulation, and the use of test accuracy (Fig. 4) as an independent performance metric is a positive feature. However, the empirical evidence is currently insufficient: the central comparative claim relies on simulations without error bars, seed counts, statistical tests, or hyperparameter details, and the optimization problem omits the hard per-device time and energy constraints stated in the system model. These issues are fixable but require substantial revision.","major_comments":[{"comment":"The system model states that \"the total time and energy consumed by each device during model training must not exceed its specified maximum limits,\" but Problem (21) contains no constraints of the form T_cmp^n + T_com^n <= T_max^n or E_cmp^n + E_com^n <= E_max^n. The objective only includes soft penalties V_T^n and V_E^n, and since T in Eq. (15) is the global maximum over all devices, the penalty for device n does not even equal its own time consumption. The PDQN can therefore select actions that violate the stated per-device budgets, and the reported VoL advantage over baselines may be an artifact of infeasible schedules. Please add explicit hard constraints, or justify why the soft penalties suffice and report the fraction of actions that violate the stated limits.","section":"Section II-C and Eq. (21)"},{"comment":"The requirement factor of TLW is epsilon_req^n = 1 / (lambda1 * T_max^n + lambda2 * E_max^n - lambda3 * A_req^n). With the simulation ranges T_max in [0.1,10] s, E_max in [0.01,1] J, and A_req in [0.7,1.0], the denominator can be negative or zero for reasonable weight choices, making epsilon_req^n negative and inverting the intended task-priority ordering. The paper does not specify the weight values, nor does it impose any positivity constraint or normalization. This is load-bearing because the objective (21a) maximizes a sum of TLW-weighted values, and negative TLW would make the objective no longer represent a meaningful value. Please specify the weights, constrain or normalize them to keep the denominator positive, and provide a sensitivity analysis.","section":"Section II-D, Eq. (17)"},{"comment":"The central claim that the proposed scheme \"significantly outperforms\" baselines is not supported by the reported experimental evidence. The figures appear to show single runs: there are no error bars, no number of random seeds, no statistical tests, and no table of DRL hyperparameters (learning rates, mini-batch size M, buffer capacity G, discount factor kappa, soft-update coefficient zeta, exploration noise schedule, or network architecture). The baseline set is also limited to DDPG, OMA, EW, and RRA, with no comparison to an optimal or exhaustive solver on small instances. To substantiate the comparative claim, please report mean +/- standard deviation over multiple seeds, provide all hyperparameter values, run significance tests, and discuss whether the accuracy levels in Fig. 4 are consistent with typical CIFAR-10 results for the chosen architecture.","section":"Section IV, Figs. 3-5"},{"comment":"Using the objective (21a) directly as the DRL reward (Eq. (22)) means that the VoL curves in Fig. 5 largely restate the very function being optimized. The independent evidence for learning quality is Fig. 4, which reports test accuracy, but the meta-learning setup is not described in enough detail: no model architecture, no local update steps, no meta-learning rates alpha and beta, and no fine-tuning procedure. Without these details, it is hard to judge whether the absolute accuracy levels are meaningful or whether the comparison to the OMA and EW schemes is apples-to-apples. Please document the complete FML training configuration and report the VoL results alongside actual resource-feasibility statistics.","section":"Section III-A3 and Fig. 4"}],"minor_comments":[{"comment":"In the state-space description, the text says \"We define the agent's action space from two aspects\" but then lists the state components h_n and epsilon_n. This appears to be a typo: it should say \"state space.\"","section":"Section III-A1"},{"comment":"The target Q-value formula uses both S_m and \\tilde{S}_m with inconsistent notation; the next state should be S_{m+1} consistently, and the max over discrete actions should be made explicit. Please correct the formula.","section":"Eq. (26)"},{"comment":"The indicator function notation in Eq. (13) is not defined clearly; the symbol \"/BD\" is confusing and should be replaced with a standard indicator notation such as \\mathbb{1}{...}.","section":"Eq. (13)"},{"comment":"In Algorithm 1, line 9 says \"if reach G\"; this should be \"if the replay buffer is full\" or \"if buffer size reaches G,\" and the update condition should be specified precisely.","section":"Algorithm 1"},{"comment":"In the conclusion, \"PQDN\" is a typo and should be \"PDQN.\" Also, the final sentence says \"simulation results demonstrated\" but no error bars or statistical tests are given, so \"demonstrate\" overstates the evidence.","section":"Section V"},{"comment":"Several simulation parameters that are essential for reproducibility are missing, including the number of devices' data samples, the non-IID distribution parameters, the number of local SGD steps, and the number of global rounds used in Fig. 4. Please add them.","section":"Table I"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the omitted hard per-device time and energy constraints is valid and should be the primary focus of revision. The conceptual framework is coherent and the topic is within the journal's scope, but the empirical validation is currently not up to the standard required for the strong comparative claims made in the abstract and conclusions. I would not reject the paper, because the issues are addressable by adding constraints, providing full experimental details, and running proper statistical comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a look if you work on DRL-based scheduling for wireless federated learning. What's actually new: the Value of Learning (VoL) and Task-Level Weight (TLW) metrics in Eqs. (13)-(20), and the application of PDQN to a NOMA-based FML resource allocation problem. The MDP formulation is coherent, the hybrid discrete/continuous action space is handled cleanly, and the non-IID CIFAR-10 setup is relevant. The authors also cite the right prior work, including the JSAC paper by Yue et al. I see no overt mathematical errors.\n\nThe soft spots are real but fixable. The biggest is that Section II-C says total time and energy 'must not exceed' T_max and E_max, yet the optimization problem (21) never enforces that. The objective penalizes V_T and V_E softly, but those are not constraints, and V_T is based on the system-wide max time, not per-device time. So the agent can pick schedules that blow through a device's budget, and the reported VoL gains may be computed for infeasible solutions. The paper doesn't say whether the baselines respect the budgets either, which makes the comparison unfair regardless of direction. That's a load-bearing problem for the 'significantly outperforms' claim, and it needs to be addressed head-on: either add hard constraints to (21) or explicitly justify why soft penalties are acceptable and show the resulting schedules still respect the stated limits.\n\nSecond, the empirical evaluation is thin. No error bars, no statistical tests, no hyperparameter values for the PDQN or the baseline DDPG. The baseline set is weak: DDPG with rounding, OMA, RRA, and equal-weight rotation are not strong contemporary baselines. A comparison to a recent FML-over-wireless optimizer, like the JSAC one, would be more informative. Third, the reward is the VoL objective itself, so reporting VoL improvement is partly circular. That said, the test accuracy in Fig. 4 is an independent metric and it does favor the TLW scheme, so the circularity burden is mild.\n\nWho is this for? Researchers in wireless FL/FML resource allocation who care about task-aware scheduling. It's a conference-level paper with a potentially useful metric, but the constraint omission and evaluation gaps need revision before the central claim is credible. I'd send it to peer review rather than desk-reject it: the flaws are serious but not fatal, and the core idea deserves referee scrutiny.","headline":"A coherent, workmanlike FML-over-NOMA scheduling paper with genuinely new VoL/TLW metrics, but the optimization drops the hard per-device time/energy budgets stated as mandatory, so the headline gains may rest on infeasible schedules.","tokens_in":783,"tokens_out":861,"would_cite":false,"duration_ms":28667,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A value-of-learning metric with a task-level fairness weight, solved by a parameterized deep Q-network, lets a NOMA-based federated meta-learning system beat uniform, OMA, DDPG, and random baselines.","keywords":["federated meta-learning","non-orthogonal multiple access","value of learning","task-level weight","parameterized deep Q-network","device scheduling","resource allocation","task-oriented learning"],"falsifier":"Run the same FML simulation with task requirements drawn from real application traces instead of uniform ranges and with the VoL/TLW weights swept over a grid; if a baseline that ignores VoL/TLW matches or beats the proposed policy on test accuracy and total VoL for any plausible weight setting, the claimed advantage is an artifact of the chosen parameters rather than a property of the framework.","tokens_in":8926,"feed_emoji":"🎯","tokens_out":7047,"duration_ms":60671,"temperature":0.7,"pith_summary":"This paper tries to show that a wireless federated meta-learning system should schedule devices and allocate power and computing resources according to each device's individual task needs, not uniformly. It introduces a metric called value of learning (VoL) that rewards reaching a device's required accuracy and penalizes time and energy spent, and a task-level weight (TLW) that combines task requirements with a fairness term based on how long ago each device last contributed. The resulting joint scheduling and resource-allocation problem is a non-convex mixed-integer program, and the paper proposes a parameterized deep Q-network (PDQN) to solve it with hybrid discrete and continuous actions. Simulations over a non-IID CIFAR-10 setup show the proposed scheme reaching higher test accuracy and higher total VoL than DDPG, OMA, equal-weight, and random-allocation baselines. If correct, this provides a concrete way to make federated meta-learning task-aware in resource-constrained wireless networks.","feed_headline":"Task-aware scheduling lifts federated meta-learning accuracy","feed_subtitle":"A hybrid deep RL policy weights devices by accuracy need, time budget, and energy cost, beating uniform and random baselines.","key_machinery":"The machinery is the VoL-TLW objective together with the PDQN solver. VoL, written as $V_n = \\varepsilon_n z_n (\\eta_1 V^A_n - \\eta_2 V^T_n - \\eta_3 V^E_n)$, turns a device's accuracy requirement and its time and energy budgets into a single scalar that the scheduler maximizes. TLW, $\\varepsilon_n = 1/(\\lambda_1 T^{\\max}_n + \\lambda_2 E^{\\max}_n - \\lambda_3 A^{\\mathrm{req}}_n) + a^n_t/\\sum_i a^i_t$, combines task importance with update staleness so devices are not starved. PDQN combines a parameterized actor network for continuous power and frequency variables with a Q-actor network for the binary scheduling decisions, which is what lets one policy handle the hybrid action space of the MINLP.","core_discovery":"The central claim is that maximizing the sum of TLW-weighted VoL across devices is the right objective for task-oriented federated meta-learning, and that the PDQN algorithm can effectively optimize it. VoL for each device is defined as eta1 times the ratio of achieved to required accuracy (capped at 1) minus eta2 times the ratio of the global round time to the device's maximum tolerable time minus eta3 times the ratio of the device's energy consumption to its maximum tolerable energy. TLW weights each device by the inverse of a weighted combination of its time and energy budgets minus its required accuracy, plus an age-of-update fairness term. The paper argues that this formulation captures individual requirements better than the common-model FL objective, and its simulations indicate that the PDQN-based policy over NOMA outperforms DDPG, OMA, equal-weight scheduling, and random resource allocation in both FML accuracy and accumulated VoL.","pith_inferences":["The paper's VoL definition assumes achieved accuracy, time, and energy fully capture a task's value; for tasks where data distribution or risk profile matters, the metric would need extra terms.","One immediate testable extension is to learn the weights eta and lambda from data rather than fixing them, so the scheduler can adapt to unknown task utilities.","Another extension is to replace the uniform ranges for required accuracy, time, and energy budgets with traces from real applications; the claimed gains would be stronger if they survive that replacement.","The framework's reward clamps non-positive VoL to zero, which may encourage the agent to ignore low-value devices entirely; a different clamping or regularization could change scheduling behavior."],"forward_implications":["If VoL is the right objective, devices with high required accuracy and tight time and energy budgets should be scheduled before devices with larger budgets, even when the latter have better channels.","Because the reward function is the sum of TLW-weighted VoL, any improvement in NOMA's spectral efficiency directly raises the objective by lowering transmission time and energy penalties.","The age-of-update fairness term ensures that devices with low task importance are not permanently starved, since their TLW grows as their last update ages.","The PDQN solver, which handles both discrete and continuous actions, is the reason the proposed scheme outperforms DDPG, which must round continuous actions and loses accuracy.","The same TLW-based VoL objective can be applied to other wireless multiple-access schemes; the OMA benchmark already uses PDQN and ranks second in VoL."],"supporting_citations":[{"why":"Introduces the MAML-based federated meta-learning approach that the paper's training model extends.","marker":"[4]"},{"why":"Jointly optimizes device scheduling and resource allocation for FML in wireless networks, the baseline problem this paper re-frames with VoL/TLW.","marker":"[7]"},{"why":"Supplies the unbiased gradient estimates and personalized federated learning guarantees used in the local update equations.","marker":"[11]"},{"why":"Justifies neglecting downlink transmission due to the server's higher transmit power.","marker":"[12]"},{"why":"Provides the effective capacitance coefficient model for computation energy consumption.","marker":"[13]"},{"why":"Provides the NOMA/SIC model and decoding order used in the communication rate expression.","marker":"[14]"},{"why":"Motivates the age-of-update fairness term by showing stale local models degrade federated learning performance.","marker":"[15]"},{"why":"Supplies the parameterized deep Q-network method the paper adapts to hybrid discrete and continuous action spaces.","marker":"[16]"}],"fun_headline_variants":["Value-of-learning metric guides smarter federated meta-learning","Task-level weights improve learning value in FML over NOMA","Deep Q-network optimizes task-aware federated meta-learning","VoL-aware scheduling lifts federated meta-learning performance","Fair task weighting maximizes learning value in FML"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the simulation's uniform ranges for required accuracy, time and energy budgets, and the hand-chosen weights in the VoL and TLW formulas, adequately represent realistic diverse tasks; if they do not, optimizing the reported VoL may not reflect true task value.","fun_headline_variants_meta":{"raw":{"variants":["Value-of-learning metric guides smarter federated meta-learning","Task-level weights improve learning value in FML over NOMA","Deep Q-network optimizes task-aware federated meta-learning","VoL-aware scheduling lifts federated meta-learning performance","Fair task weighting maximizes learning value in FML"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000522,"raw_usage":{"total_tokens":2531,"prompt_tokens":956,"completion_tokens":1575,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":1497}},"tokens_in":572,"tokens_out":1575,"duration_ms":11613,"temperature":1.0,"reasoning_tokens":1497,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:52:17.551417+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same FML simulation with task requirements drawn from real application traces instead of uniform ranges and with the VoL/TLW weights swept over a grid; if a baseline that ignores VoL/TLW matches or beats the proposed policy on test accuracy and total VoL for any plausible weight setting, the claimed advantage is an artifact of the chosen parameters rather than a property of the framework.","supporting_citations":[{"cited_title":"E fficient federated meta-learning over multi-access wireless netwo rks,","cited_arxiv_id":null,"evidence_quote":"Jointly optimizes device scheduling and resource allocation for FML in wireless networks, the baseline problem this paper re-frames with VoL/TLW."},{"cited_title":"Personalized federated learning with theoretical guarantees: A model-agnostic me ta-learning approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the unbiased gradient estimates and personalized federated learning guarantees used in the local update equations."},{"cited_title":"Federated learning over wireless n etworks: Convergence analysis and resource allocation,","cited_arxiv_id":null,"evidence_quote":"Justifies neglecting downlink transmission due to the server's higher transmit power."},{"cited_title":"A survey on non-orthogonal multiple access for 5 G networks: Research challenges and future trends,","cited_arxiv_id":null,"evidence_quote":"Provides the NOMA/SIC model and decoding order used in the communication rate expression."},{"cited_title":"A PDDQNLP algorithm for energy efficient computation offloadi ng in UA V-assisted MEC,","cited_arxiv_id":null,"evidence_quote":"Supplies the parameterized deep Q-network method the paper adapts to hybrid discrete and continuous action spaces."}],"review_version":1}