{"id":"3c6af67d-5521-4dfe-8a2c-1c29bf81d269","arxiv_id":"2501.15305","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A DQN-based UAV scheduler can extend the simulated lifetime of heterogeneous edge devices under power and communication outages and prioritize devices on evacuation routes.","lead":"This paper trains a deep Q-network to schedule a UAV visiting battery-limited edge devices during simulated disasters, with some devices lacking power, some lacking communication, and data expiring if not collected. It reports that the learned policy extends network lifetime and, in rural and urban evacuation simulations, prioritizes devices on high-traffic roads.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that the UAV learns to prioritize high-traffic devices is not yet supported: Table 7 has no control policy, so the failure counts may reflect random power/communication assignments or device asymmetries rather than learned behavior.","rationale":"The reader's weakest_assumption focuses on simulator-to-reality transfer and post hoc reward selection; my concern is more local: even inside the simulator, the evacuation result has no control policy, so the causal claim that the RL agent learned to prioritize high-traffic devices is not established. The reader's rationale does list 'absence of simple baselines' among the fixable issues, so there is partial agreement, but the reader did not elevate this to the weakest assumption. I chose UNCHANGED because this is precisely the kind of missing comparison that a conditional acceptance should require: the verdict stays conditional, with the baseline experiment as an explicit acceptance condition. I am not disputing the internal consistency of the simulation code or the plausibility of the system; the concern is specifically that the quantitative support for the central 'learned prioritization' claim is incomplete. A single controlled re-run of the existing test suite with random and greedy schedulers would settle whether the observed failure pattern is attributable to the learned policy or to scenario confounds.","tokens_in":13382,"tokens_out":4388,"duration_ms":43767,"concrete_test":"Re-run the 30 rural and 30 urban evacuation tests under three schedulers on the same random seeds and device configurations: (1) uniform random action selection, (2) a simple greedy policy such as visiting the device with the lowest remaining battery fraction, and (3) the trained DQN. Tabulate per-device first-failure frequencies exactly as in Table 7. If the random or greedy controls reproduce the observed pattern, the learned-prioritization conclusion is not supported; if the DQN distribution differs significantly by a permutation test or two-proportion test with p < 0.05 after multiple-comparison correction, the claim would be materially strengthened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline outcome in the evacuation study (Section 5.1) is that, across 30 randomized tests per scenario, devices near high-traffic roads were 'prevented from failure' (Table 7). This is the central evidence that the trained DQN policy produces the claimed operational benefit. However, every result is generated by the trained DQN; no random-action, round-robin, nearest-device, or greedy baseline is reported for the same test suite. The per-test power/communication configurations are randomized, and the fixed device placements have heterogeneous batteries, tasks, and processing rates. A first-failure distribution concentrated on low-traffic devices could therefore be produced by (i) the reward shaping that explicitly adds bonuses for high-density roads, (ii) incidental correlations between road proximity and which devices are left unpowered or unconnected, or (iii) device energy and processing asymmetries, even with an untrained or random scheduler. The anomalies the authors themselves note (device 8 in the rural scenario never fails despite low traffic; device 11 near low density fails often) indicate the pattern is not clean. Without a control policy, the sentence 'the UAV successfully learns to prioritize devices close to high traffic flow' is unsupported, and the paper's strongest claim rests on a missing comparison rather than a demonstrated policy effect.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a UAV-assisted mobile edge computing system for disaster scenarios in which heterogeneous edge devices (e.g., Raspberry Pi, Jetson Nano) may lose power supply or communication connectivity. The authors formulate an optimization problem over UAV routing that trades off device battery consumption against the age of uncommunicated data, and then train a deep Q-network (DQN) to select which device to visit in each time slot. The system is evaluated in synthetic scenarios with varying numbers of powered and connected devices, and in two SUMO-based evacuation case studies (rural Round Lake and urban Albany). The headline claims are that the learned policy extends the operational lifetime of the network and identifies the device most likely to fail first, and that in the evacuation scenarios it learns to prioritize devices near high-traffic roads.","tokens_in":13686,"tokens_out":4494,"duration_ms":39154,"significance":"If the results are correct, the paper would make a modest contribution: it adds power and communication outage constraints and data-age constraints to a well-studied UAV-assisted MEC problem, and demonstrates a DQN-based scheduler in realistic traffic-evacuation settings. The use of established hardware parameters from prior work (Tables 2 and 3) and the connection to SUMO traffic simulation are strengths. However, the formal problem formulation contains material errors, and the central behavioral claim about traffic-aware prioritization is not supported by the experimental design because no baseline policy is tested. The reward function is also selected on the same episode-length metric that constitutes the main evaluation, which weakens the persuasiveness of the reported lifetime improvements.","major_comments":[{"comment":"The claim that \"the UAV successfully learns to prioritize devices close to high traffic flow\" is not supported by the experiments as presented. Table 7 reports failure counts for the trained DQN only; there is no random-action, round-robin, nearest-device, or greedy baseline under the same 30 random power/communication configurations. The observed pattern could arise from the explicit traffic-density bonus in the reward function, from incidental correlations between device location and which devices are assigned outages, or from asymmetries in battery capacity and processing rates. The authors should add control policies and report the same failure-count table for them; this is necessary to attribute the prioritization to the learned policy.","section":"Section 5.1, Table 7"}],"minor_comments":[{"comment":"The title contains a typo: \"UA V-Assisted\" should be \"UAV-Assisted\".","section":"Title"},{"comment":"The sentence \"since the the different optimization problems are non-convex\" has a duplicated article; it should read \"since the different optimization problems are non-convex\".","section":"Section 2"},{"comment":"The description of the reward function is ambiguous. The text first gives a generic example r_t = log(A_t)+log(U_t)+log(M_t)-log(O_t), then states that the chosen reward is \"the sum of the number of time slots elapsed and the log of the ratio of the age of data in the chosen device to the oldest data in any device.\" The reader must infer that this is U + log(A/O); please state the exact functional form in a display equation and define all symbols consistently.","section":"Section 3.2"},{"comment":"The noise power is listed as -100 dBm in Table 1, but in Eq. (4) it is denoted sigma^2. dBm is a unit of power, not power squared; please clarify the units and whether sigma^2 is the noise power or the noise variance.","section":"Section 4, Eq. (4)"},{"comment":"Table 5 reports mean episode lengths without any measure of variance, and Figure 3b shows that initial conditions strongly affect the maximum possible episode length. Reporting standard deviations or confidence intervals across seeds would strengthen the quantitative claims.","section":"Section 5"},{"comment":"The explanation of the traffic-density reward is brief: it says the RL system was \"programmed to receive additional rewards if an edge device was situated near a high-density road, in proportion to the observed average density.\" Please specify how the density is computed, how the proportionality is calibrated, and whether this shaping is included in the final reward used for the results in Table 7.","section":"Section 5.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses a timely application area, but the formal model errors and the absence of baselines in the evacuation study are substantial. The authors do not appear to release code, which makes it difficult to verify whether the simulator matches the corrected equations. The paper's central claim about learned traffic-aware prioritization is currently an overstatement relative to the evidence. I believe the work is salvageable with a major revision that corrects the formulation and adds control experiments, but it is not acceptable in its present form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — this is a solid simulation study of using DQN to schedule a single UAV serving 12 heterogeneous edge devices that lose power and/or connectivity during a disaster, with data-age expiration as an additional constraint. The problem is a sensible extension of the UAV-MEC literature, and the evacuation scenarios (rural Round Lake, urban Albany) are genuinely new context. The paper is also re-implementable: parameters come from published sources, the Gym environment is described, and Table 6 gives a useful comparison of reward functions and RL methods.\n\nThe best thing in the paper is the careful experimental setup. They vary the number of powered and communicating devices, report episode lengths for each combination, and test multiple reward functions. The observation that the log-combined reward works best across scenarios is plausible. The identification of the most-failure-prone device as a maintenance priority is a practical idea.\n\nThe soft spots are real, though none are disqualifying. Equation (1d) has a sign error: when a device is served it subtracts E_offload minus E_process, which can increase battery rather than decrease it for typical values. That needs fixing. Equation (1f) never resets data age for communicating devices that are not visited; if a device can communicate, its data age should reset regardless of UAV visits. The reward function was chosen after the fact from Table 6 to maximize the very episode-length metric the paper reports, so the headline lifetimes partly reflect reward selection, not an independent policy effect. That's not fatal, but the paper should acknowledge it.\n\nThe more serious concern, flagged by the stress-test, is Table 7. The claim that the UAV 'learns to prioritize devices close to high traffic flow' has no control policy. No random-action, greedy, or round-robin baseline is reported for the same 30 random configurations. The failure frequencies could be driven by the reward shaping (which adds explicit bonuses for high-density roads), by random power/communication assignments that happen to correlate with road proximity, or by device asymmetries. The authors openly note anomalies (rural device 8 never fails despite low traffic; device 11 fails often). Without a baseline, the learned-prioritization claim is unsupported.\n\nAlso, there are no error bars or confidence intervals on Table 5, and no code or data release. For a simulation paper, that's a significant omission if the results are to be reproduced.\n\nOverall: worth engaging with, but as a conditionally acceptable simulation study, not a definitive demonstration. It deserves peer review with heavy revision. For a reader interested in UAV-assisted edge computing, the problem formulation and experimental structure are useful; the quantitative claims need more controls and re-analysis.","headline":"A plausible, re-implementable simulation study of DQN for UAV edge computing under disasters, but the headline claims about learning to prioritize traffic-critical devices are not yet supported by the experiments as presented.","tokens_in":14147,"tokens_out":2540,"would_cite":false,"duration_ms":20943,"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 drone-mounted scheduler learns which disaster sensors will fail first and extends their working life.","keywords":["UAV-assisted edge computing","disaster resilience","deep Q-network","reinforcement learning","task offloading","device lifetime","data age","evacuation simulation"],"falsifier":"Run the same deep Q-network training in a physical testbed or in a second, independently written simulator with the same device types and outage configurations, and compare the first-failing device and episode lengths; if the predicted critical device differs in a substantial fraction of configurations, the simulator-grounded claim fails to transfer.","tokens_in":13199,"feed_emoji":"🚁","tokens_out":3647,"duration_ms":31390,"temperature":0.7,"pith_summary":"This paper claims that a single UAV, flown by a deep Q-network policy, can stretch the operating lifetime of heterogeneous edge-computing devices during disasters where some devices lose power and others lose communication. The UAV visits devices to offload their computations and relay their data, and the learned policy balances battery drain against the age of uncommunicated data. The authors report that the policy identifies the device most likely to fail first in a given configuration, giving maintenance crews a concrete priority target. They also show, in simulated evacuations of a rural town and an urban downtown, that devices near high-traffic roads are kept from failing first even when power and connectivity outages are randomized.","feed_headline":"Drone scheduler learns which disaster sensors fail first","feed_subtitle":"Reinforcement learning lets a drone extend device lifetimes during outages and flags the most critical device.","key_machinery":"The machinery is a deep Q-network scheduler whose state is a vector of remaining battery fractions and data ages for all devices, whose action is choosing which device the UAV visits next, and whose reward adds UAV time steps, the log ratio of chosen-device data age to oldest data age, and battery and age terms. The environment is a Gym simulation where each time slot consists of the UAV flying to a device, offloading computation, and all other devices processing locally, with Shannon's formula giving the transmission rate. In the evacuation cases, the reward is augmented in proportion to the traffic density near each device, which is what steers priority toward high-flow roads.","core_discovery":"In the paper's own terms, the discovery is that deep Q-network reinforcement learning can learn an effective UAV scheduling policy for a disaster-stricken edge network without a model of the disaster dynamics, under combined power and communication failures. The system maximizes the number of time slots before any device fails, where failure is defined as battery exhaustion or data age exceeding a limit, and in doing so it identifies the most critical device. In the evacuation simulations, the learned policy reliably protects devices on high-density evacuation routes: none of the rural high-traffic devices failed first in 30 randomized tests, and the urban high-traffic devices were likewise shielded despite occasional anomalies.","pith_inferences":["Beyond the paper's claims, the reward using the ratio of selected-device data age to oldest data age suggests a broader principle: prioritizing the device with the highest urgency gap, a rule that could transfer to other scheduling domains.","An implicit testable extension is to calibrate the imported channel and hardware parameters with real UAV-device measurements and re-run the same simulator; large deviations in per-device lifetime would indicate where the model departs from reality.","The anomalies in the evacuation results hint that the policy's protection is statistical rather than guaranteed, and confidence intervals over many random configurations would characterize that uncertainty more precisely."],"forward_implications":["If the policy works as modeled, operators could use the first-failing device prediction to dispatch maintenance to the single most critical device during an outage.","The episode-length results give a quantitative map of how network lifetime degrades as the number of devices with power and communication availability drops from 12 to 4.","The evacuation results imply that traffic-density-weighted rewards can embed domain knowledge into the scheduler without changing the reinforcement learning algorithm.","Since the state is just battery fractions and data ages, the same deep Q-network approach can be resized to any number of devices by enlarging the state vector.","Multiple-UAV deployment, flagged as future work, is a natural next step because single-UAV lifetime is ultimately capped by the UAV battery."],"supporting_citations":[{"why":"Supplies the per-device power consumption and processing-rate tables for the five device types across the vision tasks used in the simulator.","marker":"[24]"},{"why":"Supplies the communication and channel model parameters used to compute transmission rates.","marker":"[3]"},{"why":"Establishes the UAV mobile edge computing and deep Q-network trajectory-control approach that this paper extends.","marker":"[28]"},{"why":"Motivates the use of deep reinforcement learning for UAV path planning in mobile edge computing.","marker":"[12]"},{"why":"Provides Shannon's capacity formula used to derive transmission time from channel gain.","marker":"[20]"},{"why":"Provides the SUMO traffic simulator used to generate evacuation traffic densities for the rural and urban scenarios.","marker":"[8]"},{"why":"Supplies the stable-baselines implementation of the deep Q-network algorithm used for training.","marker":"[17]"}],"fun_headline_variants":["Drone AI predicts which disaster sensor dies first","Reinforcement learning drones extend sensor lifetimes","UAV scheduler learns to save critical sensors in crises","Deep RL for drones shields failing edge devices","Drone policy identifies the next sensor to fail"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The policy's success depends on the simulator's dynamics, including the imported power and channel parameters and the reward function chosen from the same simulator, matching real disaster conditions closely enough that a trained policy keeps working outside the simulation.","fun_headline_variants_meta":{"raw":{"variants":["Drone AI predicts which disaster sensor dies first","Reinforcement learning drones extend sensor lifetimes","UAV scheduler learns to save critical sensors in crises","Deep RL for drones shields failing edge devices","Drone policy identifies the next sensor to fail"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000546,"raw_usage":{"total_tokens":2574,"prompt_tokens":872,"completion_tokens":1702,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":488,"completion_tokens_details":{"reasoning_tokens":1632}},"tokens_in":488,"tokens_out":1702,"duration_ms":11356,"temperature":1.0,"reasoning_tokens":1632,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:24:19.963706+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same deep Q-network training in a physical testbed or in a second, independently written simulator with the same device types and outage configurations, and compare the first-failing device and episode lengths; if the predicted critical device differs in a substantial fraction of configurations, the simulator-grounded claim fails to transfer.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the per-device power consumption and processing-rate tables for the five device types across the vision tasks used in the simulator."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the communication and channel model parameters used to compute transmission rates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the UAV mobile edge computing and deep Q-network trajectory-control approach that this paper extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates the use of deep reinforcement learning for UAV path planning in mobile edge computing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides Shannon's capacity formula used to derive transmission time from channel gain."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the SUMO traffic simulator used to generate evacuation traffic densities for the rural and urban scenarios."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the stable-baselines implementation of the deep Q-network algorithm used for training."}],"review_version":1}