{"id":"a2f9f263-488b-42c7-a849-521995457269","arxiv_id":"2412.00386","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A WGAN-based data augmentation and a knowledge-driven channel knowledge map are combined with PPO to improve UAV trajectory design, with reported gains that may be an artifact of circular evaluation.","lead":"This paper uses a Wasserstein GAN to generate synthetic wireless channel data, trains a knowledge-driven channel map, and then uses reinforcement learning to plan UAV flight paths. The authors report faster mission completion and higher throughput than baseline methods, but the evaluation appears to use the trained model's own predictions as the ground truth.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains may be artifacts of evaluating each agent against its own channel surrogate (Algorithm 3: L ← G); no common ground-truth channel is used for final comparison.","rationale":"The reader's weakest_assumption correctly identifies the core issue: the RL agent is evaluated against the CKM's own predicted channel losses (Algorithm 3: L ← G), so improvements in flight time and throughput presuppose that the CKM is an accurate, unbiased surrogate. My analysis confirms this is the single most load-bearing concern. The paper does not provide a common ground-truth evaluation for all baselines, does not report error bars or confidence intervals, and does not supply code or sufficient simulation parameters to reproduce the results. These deficiencies compound the circular evaluation: even if the CKM is reasonably accurate, the comparison as presented could still be dominated by surrogate mismatch rather than by the proposed framework's merits. A single concrete check—evaluate all learned policies against a fixed, held-out channel simulator—would settle whether the reported gains are real. Because that check has not been performed and the paper's claims as written are unsupported, the reader's REJECT verdict stands. I see no reason to move the verdict; the concern is the same as the reader's, and the missing evidence is decisive at this stage.","tokens_in":11436,"tokens_out":2687,"duration_ms":29756,"concrete_test":"Run a shadow evaluation: fix a ground-truth channel simulator (the same A2G/ray-tracing data source used to generate the training dataset, or a held-out set of channel attenuation values), then execute the trained KDCKM-PPO and LoS-PPO policies in that fixed simulator, computing flight time and throughput from ground-truth L rather than from each policy's internal prediction. Report both policies under the same ground truth. If KDCKM-PPO no longer beats LoS-PPO in this common setting, the central performance claim collapses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central performance claim—that KD-CKM with WGAN augmentation yields 37s flight time and 16.18 bps throughput versus 45s and 14.97 bps for LoS-PPO—is not supported because the evaluation is circular. In Algorithm 3, the RL environment obtains the channel loss as L ← G, where G is the trained CKM's prediction, and then uses that predicted loss to compute throughput and reward. The LoS-PPO baseline analogously uses the LoS probability model (Eq. 3) as its channel. Thus the two policies are optimized and scored in different virtual worlds: each sees the world through its own surrogate. The reported throughput values are not measurements of the same physical channel; they are the CKM's self-consistent belief about the channel for KDCKM-PPO and the LoS model's belief for LoS-PPO. If the CKM is optimistically biased—e.g., because WGAN augmentation reinforces simulation artifacts or because the knowledge block overfits the training distribution—the KDCKM-PPO policy can exploit these optimistic predictions to achieve lower 'flight time' and higher 'throughput' without actually performing better in the real channel. The paper provides no evaluation against a common ground-truth channel simulator or held-out real channel data, so the 37s vs 45s gap and throughput difference may be entirely an artifact of comparing two different surrogate channels rather than two policies on the same task. This is not a minor detail: the entire contribution of integrating CKM into trajectory design rests on the accuracy of the CKM as a proxy, and the paper never validates that proxy in the final trajectory evaluation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an AIGC-enhanced framework for UAV trajectory design. A WGAN is used to augment a channel-attenuation dataset, a knowledge-driven Channel Knowledge Map (KD-CKM) is trained to predict channel loss from UAV/ground-user positions and environmental features, and a PPO-based reinforcement learning agent uses the CKM predictions to plan a trajectory that minimizes flight time while satisfying user communication requirements. Experiments report improvements in CKM prediction accuracy and, in the trajectory design task, a flight time of 37 s and throughput of 16.18 bps for KDCKM-PPO versus 45 s and 14.97 bps for the LoS-PPO baseline.","tokens_in":11768,"tokens_out":3602,"duration_ms":36315,"significance":"If the reported performance gains were established, the paper would make a useful contribution to the growing literature on channel knowledge maps and AI-generated data for UAV communications. Its algorithmic pipeline is clearly structured, and the idea of embedding a physical LoS model into the CKM architecture is reasonable. However, the central performance claims are not supported by the current experimental design: the trajectory comparison evaluates each RL agent against a different channel surrogate, and the data-augmentation evaluation does not demonstrate generalization beyond the training distribution. With the reported design, the results cannot be interpreted as evidence that the proposed framework outperforms the baselines in a common physical channel.","major_comments":[{"comment":"The trajectory evaluation is circular. In Algorithm 3 the environment obtains the channel loss as L ← G, where G is the trained CKM for KDCKM-PPO and the LoS probability model for LoS-PPO. The two agents are therefore trained and scored in different virtual worlds, each using its own surrogate for the true channel. The reported flight-time and throughput differences (Table I: 37 s vs 45 s; 16.18 vs 14.97 bps) may reflect optimism or bias in the surrogates rather than differences in the learned policies. To support the central claim, the authors must evaluate both policies against a common ground-truth channel (e.g., the simulator that generated the original dataset, or a separate ray-tracing tool) and report the resulting flight time and throughput.","section":"III-D (Algorithm 3) and Table I"},{"comment":"The WGAN is trained on the original dataset and generates samples whose distribution is 'closely aligned to that of the original data' (Section IV-A, Fig. 3). Such synthetic data carries no independent information about the true channel beyond the empirical training distribution, so it cannot be claimed to improve the CKM's generalization to new environments. The CKM accuracy comparison (Fig. 4) is performed on a validation split from the same distribution, so the reported MAPE and MSE reductions do not demonstrate that AIGC adds information. The authors should evaluate the CKM on a held-out test set from a different environment or region, or compare against simpler augmentation baselines (e.g., additive noise or interpolation).","section":"IV-A (WGAN data augmentation) and Fig. 4"},{"comment":"Key experimental details are missing, preventing replication and statistical assessment. The paper does not specify how the 'real channel attenuation data' were generated, which physical model or simulator produced the ground-truth labels, or the exact weights in the MDP reward function (Section III-C lists r1–r4 but gives no weights). Table I reports single values with no error bars, confidence intervals, or multiple random seeds. The absence of these details makes it impossible to judge whether the reported 37 s vs 45 s gap is significant or robust. The authors should provide the full setup and run multiple seeds.","section":"IV (Simulation setup) and III-C (Reward function)"}],"minor_comments":[{"comment":"The symbol 'a' is used both as a parameter in the LoS probability formula and in the exponent term within the same equation, which is confusing; please rename one of them.","section":"Section II-B, Eq. (3)"},{"comment":"The line 'Apply knowledge block in hidden layer: ˙H ← ˆL + res(x)' is unclear. The notation res(x) is not defined, and the relationship between the knowledge-block output and the residual connection should be explained.","section":"Algorithm 2"},{"comment":"The caption states 'xU, yU, zU for GUs' positions' but the figure legend refers to UAV positions; the labels are inconsistent and should be corrected.","section":"Fig. 3 caption"},{"comment":"The paper consistently writes 'UA V' with a space; the standard spelling 'UAV' should be used.","section":"Throughout"},{"comment":"The conclusion states that 'AIGC offers superior generalization,' but no generalization experiments are presented; this overclaims beyond the reported results.","section":"Section V (Conclusion)"}],"recommendation":"major_revision","confidential_remarks":"The central claim of the paper is currently unsupported because of the circular evaluation in Algorithm 3. The issue is potentially fixable by re-running the RL comparison against a common ground-truth channel simulator, but this is a substantial revision, not a minor edit. The paper is within the scope of the journal, and the proposed pipeline is interesting; I would encourage the authors to address the evaluation design and provide complete experimental details."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's central trajectory claim is not supported. In Algorithm 3, the RL environment gets channel loss from the trained CKM (L ← G), while the LoS-PPO baseline uses the LoS probability model. Each policy is therefore optimized and scored in a different virtual world. The reported 37s versus 45s flight time and 16.18 versus 14.97 bps throughput are not measurements of the same physical channel; they are self-consistent beliefs of two different surrogates. The stress-test note is correct, and this is a load-bearing flaw, not a detail. Without a common ground-truth channel simulator or held-out real data, the integrated contribution is unvalidated.\n\nThat said, the paper is not without merit. The knowledge-driven CKM architecture — embedding a physics-based LoS block into the hidden layers of a neural network — is a reasonable idea, and the CKM prediction evaluation (MSE, MAPE, radar plot) is separate from the trajectory circularity. The writing is clear, and the system description is complete enough to follow. The WGAN data augmentation is standard and, as the paper admits, generates data from the original distribution, so it adds no independent information; that part is over-claimed.\n\nMissing details (no error bars, no code, vague reward weights) are secondary to the circularity. The main soft spot is that the trajectory results are artifacts of comparing two surrogates. If the authors reran the comparison against a common channel model, the true result could be much less impressive.\n\nWho is this for? Researchers working on CKM-assisted UAV communications, particularly those interested in integrating learned channel models with RL. The CKM construction piece might be a useful reference, but the trajectory conclusions should not be cited. It deserves a serious referee because the integration is novel and the flaw is fixable in principle — but as it stands, I'd reject it. Give it a chance in review if the authors commit to a common-channel evaluation.\n\nRecommendation: send to peer review, but expect major revision or rejection based on the evaluation redesign.","headline":"The trajectory comparison is invalid because each policy is evaluated against its own channel surrogate, so the headline 37s vs 45s gain is not meaningful.","tokens_in":12250,"tokens_out":2067,"would_cite":false,"duration_ms":22181,"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 WGAN-augmented channel knowledge map lets a UAV finish its route in 37 seconds at higher throughput than LoS-only baselines.","keywords":["AIGC","WGAN","channel knowledge map","data augmentation","UAV trajectory design","reinforcement learning","PPO","channel prediction"],"falsifier":"Fly the trained KD-CKM policy in a high-fidelity ray-tracing channel simulator or against measured channel data at the same 15 user locations, and compare each user's received power and the mission completion time with the CKM's predictions; if the CKM systematically predicts lower path loss than the real channel delivers, the 37-second completion time will not be achieved.","tokens_in":11237,"feed_emoji":"📡","tokens_out":8900,"duration_ms":63525,"temperature":0.7,"pith_summary":"UAVs serving as aerial base stations need accurate channel-loss predictions to plan efficient routes, but collecting enough channel data is slow and expensive. This paper argues that a generative AI technique, a Wasserstein generative adversarial network (WGAN), can manufacture realistic channel-gain samples that enlarge the training set, and that a channel knowledge map (CKM) with a physics-based path-loss model embedded in its hidden layers can turn those samples into accurate, location-specific predictions. When those predictions are fed into a Proximal Policy Optimization (PPO) agent, the UAV completes its mission faster and achieves higher throughput than baselines that rely only on a line-of-sight probability model. In the simulated scenario, the proposed KD-CKM policy finishes in 37 seconds with 16.18 bps throughput, versus 45 seconds and 14.97 bps for LoS-PPO. The value, if the claim holds, is a recipe for making trajectory optimization practical in environments where real channel data is scarce.","feed_headline":"WGAN-fed channel maps cut UAV mission time from 45 to 37 seconds","feed_subtitle":"AI-generated channel data plus a physics-aware map makes UAV base stations faster and more efficient.","key_machinery":"The load-bearing object is the knowledge-driven channel knowledge map (KD-CKM): a neural network whose hidden layers receive a concatenation of learned features and the output of a line-of-sight probability path-loss model, so physics guides the prediction. The WGAN supplies synthetic channel-gain samples to the CKM's training set, and the trained CKM's channel-gain prediction is inserted into the reinforcement-learning environment as the throughput and reward calculation (Algorithm 3's L ← G). The comparison baselines are LoS-PPO and LoS-BCD, which use only the analytical line-of-sight model without site-specific learning.","core_discovery":"The central claim is that AIGC, applied as WGAN-based data augmentation plus a knowledge-driven CKM, improves both channel-loss prediction accuracy and the quality of UAV trajectories planned from those predictions. The KD-CKM architecture embeds a line-of-sight probability path-loss model as a knowledge block inside the hidden layers of a residual network, so the network combines data-driven features with physics-based attenuation estimates. Trained on original and WGAN-generated samples, this CKM predicts channel gain from UAV and user positions and environment features, and the PPO agent's reward and throughput calculation uses those predictions. In the reported simulation, the integrated system reduces flight time from 45 to 37 seconds and raises throughput from 14.97 to 16.18 bps relative to a LoS-probability PPO baseline, and from 81 to 37 seconds relative to a block-coordinate-descent baseline.","pith_inferences":["A testable extension is to close the evaluation loop: train the same PPO policy on the CKM's predictions, then fly it against a ray-tracing simulator or measured channel data; the gap between predicted and actual performance would quantify how much of the 37-second gain is genuine.","If the CKM is optimistic, the reported trajectory gains could shrink or vanish in deployment, because the agent's reward and throughput are computed from the map's own channel gains.","The same WGAN-plus-knowledge-block recipe could transfer to other site-specific radio maps, such as indoor or street-level maps, if the embedded knowledge block is replaced with an appropriate propagation model for that environment."],"forward_implications":["In the simulated urban scenario, KD-CKM-PPO completes the mission in 37 seconds versus 45 seconds for LoS-PPO and 81 seconds for LoS-BCD.","System throughput rises to 16.18 bps with KD-CKM-PPO, compared with 14.97 bps for LoS-PPO and 7.16 bps for LoS-BCD.","WGAN data augmentation improves prediction accuracy for all tested CKM variants; the largest MSE reduction, 30.23%, occurs for the basic CKM that starts with the lowest accuracy.","Embedding the path-loss model as a hidden-layer knowledge block speeds training convergence and yields lower MAPE than using it as a plain input feature.","Because the AI-based planner performs a global search, it avoids the local-optima and convergence failures that the BCD baseline exhibits under tight initial constraints."],"supporting_citations":[{"why":"Establishes radio-map-based 3D path planning for cellular-connected UAVs, the trajectory-design setting this paper extends.","marker":"[1]"},{"why":"Supplies the deep-reinforcement-learning approach for UAV trajectory design and resource allocation that the PPO agent is built on.","marker":"[2]"},{"why":"Frames the data-requirement question for channel knowledge map construction, motivating the need for augmentation.","marker":"[3]"},{"why":"Describes the environment-aware CKM framework and typical CKM-assisted communication scenarios used as the conceptual basis.","marker":"[4]"},{"why":"Provides a learned radio-map path-loss estimator (RadioUNet) as a related learned-CKM baseline.","marker":"[6]"},{"why":"Applies a CKM to compensate positioning errors in UAV missions, grounding the CKM-for-UAV use case.","marker":"[9]"},{"why":"Gives the improved Wasserstein GAN training procedure used to generate synthetic channel data.","marker":"[13]"}],"fun_headline_variants":["AI-generated channel maps make UAV routes 18% faster","Physics-aware AI cuts UAV mission time by 8 seconds","WGAN-trained CKM shrinks UAV flight time by 18%","AI + physics yields faster, higher-throughput UAV links"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported flight-time and throughput gains assume the channel knowledge map's predicted losses are an accurate, unbiased stand-in for the true wireless channel, because the reinforcement-learning agent is trained and scored against the map's own predictions rather than against real channel measurements.","fun_headline_variants_meta":{"raw":{"variants":["AI-generated channel maps make UAV routes 18% faster","Physics-aware AI cuts UAV mission time by 8 seconds","WGAN-trained CKM shrinks UAV flight time by 18%","AI + physics yields faster, higher-throughput UAV links"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000581,"raw_usage":{"total_tokens":2683,"prompt_tokens":842,"completion_tokens":1841,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":1772}},"tokens_in":458,"tokens_out":1841,"duration_ms":14139,"temperature":1.0,"reasoning_tokens":1772,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:26:11.763083+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fly the trained KD-CKM policy in a high-fidelity ray-tracing channel simulator or against measured channel data at the same 15 user locations, and compare each user's received power and the mission completion time with the CKM's predictions; if the CKM systematically predicts lower path loss than the real channel delivers, the 37-second completion time will not be achieved.","supporting_citations":[{"cited_title":"Radio Map-Based 3D Path Planning for Cellular-Connected UA V ,","cited_arxiv_id":null,"evidence_quote":"Establishes radio-map-based 3D path planning for cellular-connected UAVs, the trajectory-design setting this paper extends."},{"cited_title":"Deep reinforce- ment learning based trajectory design and resource allocation for UA V-assisted communications,","cited_arxiv_id":null,"evidence_quote":"Supplies the deep-reinforcement-learning approach for UAV trajectory design and resource allocation that the PPO agent is built on."},{"cited_title":"How much data is needed for channel knowledge map construction?,","cited_arxiv_id":null,"evidence_quote":"Frames the data-requirement question for channel knowledge map construction, motivating the need for augmentation."},{"cited_title":"A Tutorial on Environment-Aware Communi- cations via Channel Knowledge Map for 6G,","cited_arxiv_id":null,"evidence_quote":"Describes the environment-aware CKM framework and typical CKM-assisted communication scenarios used as the conceptual basis."},{"cited_title":"RadioUNet: Fast Radio Map Estimation With Convolutional Neural Networks,","cited_arxiv_id":null,"evidence_quote":"Provides a learned radio-map path-loss estimator (RadioUNet) as a related learned-CKM baseline."},{"cited_title":"Positioning error compensation by channel knowledge map in UA V communication missions,","cited_arxiv_id":null,"evidence_quote":"Applies a CKM to compensate positioning errors in UAV missions, grounding the CKM-for-UAV use case."},{"cited_title":"Improved training of Wasserstein GANs,","cited_arxiv_id":null,"evidence_quote":"Gives the improved Wasserstein GAN training procedure used to generate synthetic channel data."}],"review_version":1}