{"id":"7ecfe15c-302b-485d-b322-c553d0ece67d","arxiv_id":"2411.13104","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A deep reinforcement learning agent (MPDQN) selects RRI and transmission power in C-V2X Mode 4 to jointly minimize Age of Information and energy consumption, outperforming genetic and random baselines in simulation.","lead":"This paper uses deep reinforcement learning to pick transmission power and resource reservation intervals in C-V2X vehicle networks, aiming to reduce both data staleness (Age of Information) and energy use. Simulations with multi-priority message queues and NOMA decoding suggest the learned policy beats genetic and random baselines on both metrics.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Receiver AoI update in Eq. (15) contradicts the accompanying text (increment by 1 vs by RRI Γ); the objective and reported metric depend on this, so the claimed AoI superiority is not yet established.","rationale":"The reader's verdict of CONDITIONAL is reasonable. My focus is on a more direct internal inconsistency than the one highlighted by the reader: the AoI update on failed transmissions is specified in two mutually incompatible ways within the same subsection. Unlike the Eq. (19) concern, which depends on how an external model is imported and whether NOMA changes the meaning of 'collision,' this is a self-contradiction in the paper's own model. Since AoI is both the optimized quantity and the reported performance metric, either implementation error materially changes the results. The released code makes this checkable, and the paper deserves credit for releasing it; the fix is likely straightforward but the numerical claims should be regenerated. I do not see grounds to reject outright, since the relative comparison could survive the correction, but the central claim as written is not yet supported. The reader's weakest_assumption on Eq. (19) is related but different; I partially agree because both are model-validity issues, but the internal contradiction is more decisive and more concrete.","tokens_in":22485,"tokens_out":10089,"duration_ms":99230,"concrete_test":"Inspect the released GitHub repository (qiongwu86/DRL-Based-Optimization-for-Information-of-Age-and-Energy-Consumption-in-C-V2X-Enabled-IoV) at the function updating Φ_{i→j}^t when u=0; determine whether the increment is 1 or Γ. Then rerun the experiments of Fig. 9(a) with the opposite update rule and compare the MPDQN/GA/random AoI curves. If the ranking of algorithms or the margin of MPDQN over GA changes materially (e.g., the gap falls below the typical variation between random seeds), the claimed superiority is an artifact of the chosen (and unstated) update rule.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section III-E, Eq. (15) defines the receiver AoI update on transmission failure as Φ_{t+1} = Φ_t + 1, yet the following paragraph states that because C-V2X Mode 4 transmissions occur on reserved resources, a failed transmission forces a wait equal to the RRI size, 'resulting in an increase in Φ by Γ' (Γ ∈ {20,50,100}). These two updates differ by up to two orders of magnitude. The AoI appears in the optimization objective Eq. (25) and the reward Eq. (31), and is the headline metric in Figs. 6, 9, and 10. If the code implements the +1 update, AoI values are substantially underestimated for large Γ, biasing the comparison between RRI choices and inflating the apparent benefit of the learned policy; if it implements the +Γ update, Eq. (15) is simply wrong and the model description is not reproducible. Either way, the quantitative AoI results—and the central claim of outperforming baselines—are not robust until this inconsistency is resolved. This is an internal model contradiction, independent of the external validity of the collision model in Eq. (19).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a C-V2X Mode 4 vehicular network with multiple message classes, multi-priority FIFO queues, and NOMA/SIC at the receiver. It develops analytical models for queueing AoI, receiver AoI, collision probability, SINR, and energy consumption, and formulates a joint optimization of the resource reservation interval (RRI) and transmit power to minimize a weighted sum of average AoI and energy consumption. The optimization is solved with a hybrid discrete-continuous action DRL algorithm (MPDQN), and simulation comparisons against GA-based and random policies are presented for various vehicle counts and message sizes.","tokens_in":22809,"tokens_out":3739,"duration_ms":39627,"significance":"The combination of multi-priority queues, NOMA/SIC, AoI, and DRL-based RRI/power control is timely and the authors release source code, which aids reproducibility. The conceptual direction—prioritizing safety-critical messages and using SIC to mitigate SPS collisions—is reasonable and potentially useful for C-V2X design. However, the paper's quantitative claims are conditional on resolving an internal contradiction in the receiver AoI update, a garbled collision-probability expression, and a physically inconsistent energy model; until these are fixed, the claimed AoI and energy benefits over baselines are not established.","major_comments":[{"comment":"The receiver AoI update is internally inconsistent. Eq. (15) states that on transmission failure Φ_{t+1} = Φ_t + 1, but the paragraph immediately following states that because a failed C-V2X Mode 4 transmission forces a wait equal to the RRI size, the increase in Φ is Γ, with Γ ∈ {20, 50, 100}. These two updates differ by up to two orders of magnitude. Since Φ appears in the objective Eq. (25), the reward Eq. (31), and is the headline metric in Figs. 6, 9, and 10, the reported AoI values and the claimed superiority over baselines are not robust until this is resolved. Please correct Eq. (15) to match the implemented update or provide the code path that confirms which update is used.","section":"Section III-E, Eq. (15)"},{"comment":"The collision probability expression is garbled and its applicability under NOMA is not established. The product term has ambiguous indexing (the rendering of the product over i from 0 to Γ−1 of (1 − π/(1−π i)) is not clearly defined), and the overall formula lacks a clean derivation in this paper. More importantly, Eqs. (20) and (21) use different interference models for collisions with and without SIC, but Eq. (19) is borrowed from [51], which does not account for NOMA/SIC changing the set of undecodable interferers. If the collision probability is used to trigger transmission failures in the simulator, the AoI improvement attributed to NOMA may be an artifact of the model. Please clarify how Eq. (19) is applied in the NOMA case and provide a concrete validation, such as a comparison of SINR distributions with and without SIC for identical collision events.","section":"Section III-F, Eq. (19)"},{"comment":"The energy consumption model has a unit consistency problem. Eq. (23) defines ε_i^t = p_i^t β_i^t, which has units of power (W) when p_i^t is in watts, not energy (J). Eq. (24) multiplies this by RC_i^0, a dimensionless count, so the resulting E_i^t is still not an energy. The energy values plotted in Figs. 9 and 10 are therefore not physically meaningful without an explicit time duration or normalization factor. This is load-bearing because energy is one of the two objectives in Eq. (25) and the reward (31). Please redefine ε_i^t as p_i^t times the transmission duration (e.g., one subframe) or clearly state that a normalized energy unit is used.","section":"Section III-G, Eqs. (23) and (24)"},{"comment":"The simulation results do not include error bars, confidence intervals, or statistical tests, and the reported curves appear to be point estimates from single runs. Since the baseline comparisons in Figs. 9 and 10 are the central evidence for the claim that MPDQN 'outperforms' GA and random policies, the absence of variance information makes it impossible to judge whether the differences are significant. Additionally, the reward weight ω_1 is selected from the training curves in Fig. 7, and the trained policy is evaluated on the same simulator used for training, so there is no out-of-sample test. Please add multiple-seed results with error bars or statistical significance measures, and describe how the reported curves are averaged.","section":"Section V-B, Figs. 4–10"}],"minor_comments":[{"comment":"The action set in the loss function is written as {20, 50, 10}; this should be {20, 50, 100} to match Eq. (25b).","section":"Section IV-B, Eq. (38)"},{"comment":"The x-axis labels in Fig. 7(b) are malformed ('0.3 0.4 0.5 0.6 .7'); they should be a proper numeric axis with values 0.3 through 0.7.","section":"Section V-B, Fig. 7(b)"},{"comment":"The resource reservation model is hard to follow: the notation m_i^t, t_rk, and the condition for using Eq. (3) versus Eq. (4) is not clearly defined, and Eq. (7) contains an ambiguous bracket expression. Please rewrite these definitions more formally.","section":"Section III-C, Eqs. (3)–(7)"},{"comment":"The multi-priority transmission rule is only explicitly written for HPD and CAM; the corresponding conditions for DENM and MHD queues are not shown, even though the text says there are four queues. Please give the general expression or the full cases.","section":"Section III-D, Eq. (12)"},{"comment":"The Poisson arrival probability P(arr_{i,n}^t = 1) = λ_n e^{−λ_n} is dimensionally unclear; if λ_n is a rate per time slot, the right-hand side should be 1 − e^{−λ_n} for the probability of at least one arrival. Please define λ_n and the time unit explicitly.","section":"Section III-D, Eq. (10)"}],"recommendation":"major_revision","confidential_remarks":"The paper is closely related to the authors' previous work in [52] (Sensors 2024), and the incremental contribution—multi-priority queues and NOMA in the C-V2X AoI setting—should be stated more explicitly relative to that work. The central problem is not the novelty but the internal inconsistency of the AoI update and the lack of validation for the collision/NOMA model; these are fixable with a careful revision and re-running of simulations. If the authors can resolve Eq. (15) and provide statistical evidence for the comparisons, the paper may become suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: this is a competent, incremental simulation paper that combines multi-priority message queues, NOMA/SIC, and MPDQN for RRI/power control in C-V2X Mode 4. The specific combination is not in the literature, and the authors released code, which is genuinely useful. But the headline AoI results are not yet trustworthy because the receiver AoI update is defined inconsistently: Eq. (15) says failed transmissions increase Φ by 1 slot, while the text immediately says the increase equals the RRI size Γ (20/50/100 ms). That is up to two orders of magnitude difference, and AoI is the objective in Eq. (25), the reward in Eq. (31), and the main metric in Figs. 6, 9, and 10. Until the authors say which version the simulator implements and fix the equation, the quantitative AoI comparisons are not reproducible.\n\nWhat the paper does well: the problem is real and well-motivated; the system model covers the SPS resource selection process in reasonable detail; the multi-priority queue analysis is a sensible extension of the authors' earlier NR-V2X work; and the code release is a plus. The simulation study is fairly broad, with Nv sweeps, message-size sweeps, and a comparison against GA and random baselines.\n\nWhere it is soft: besides the AoI inconsistency, Eq. (19) for collision probability is garbled to the point of being unreadable, and it is borrowed from [51] without checking whether it remains valid under NOMA/SIC interference cancellation. The energy expression in Eq. (23) has unit issues—ε = p·β has units of power, not energy, unless multiplied by slot duration. The no-NOMA vs NOMA comparison is only shown for fixed RRI settings, not inside the DRL policy, so the benefit of NOMA in the learned policy is not isolated. There are no confidence intervals or repeated-seed results anywhere. And the reward weight ω1 is chosen based on which value trains best, which is a mild form of tuning to the objective.\n\nNone of these issues are fatal to the paper's overall direction, but the AoI contradiction is load-bearing: the paper's central claim is AoI superiority. The authors need to fix Eq. (15), report the actual update used in code, add error bars, and clarify the collision model. The math is otherwise straightforward.\n\nWho this is for: researchers in vehicular communications, especially those working on C-V2X resource allocation with AoI constraints. It is not a breakthrough, but it is a legitimate configuration of known techniques with code. A serious referee should see it, with a request for major revision on the above points.","headline":"A useful but rough simulation study combining multi-priority queues, NOMA, and MPDQN for C-V2X Mode 4; the headline AoI numbers are undermined by an internal contradiction in the AoI update rule.","tokens_in":23303,"tokens_out":3068,"would_cite":false,"duration_ms":28304,"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":"This paper claims that a deep reinforcement learning controller can jointly reduce Age of Information and energy consumption in C-V2X vehicle-to-vehicle communication, using multi-priority queues and NOMA with successive interference…","keywords":["Age of Information","C-V2X","NOMA","Successive Interference Cancellation","Deep Reinforcement Learning","Multi-priority queue","Semi-Persistent Scheduling","Vehicle-to-Vehicle communication"],"falsifier":"Run a packet-level simulation of C-V2X Mode 4 with SIC-based NOMA and compare the measured fraction of collided resource reservations against the value predicted by Eq. (19) across the same vehicle densities and RRI values; a systematic discrepancy would invalidate the collision model on which the AoI and energy improvements rest.","tokens_in":22327,"feed_emoji":"🚗","tokens_out":4749,"duration_ms":45837,"temperature":0.7,"pith_summary":"In C-V2X Mode 4, vehicles autonomously reserve radio resources, which causes collisions and degrades the freshness of safety-critical messages. This paper argues that adding multi-priority message queues and SIC-based NOMA, then letting a roadside unit learn each vehicle's transmission interval and power with a deep reinforcement learning algorithm, can cut both the average Age of Information and the average energy consumption. The proposed MPDQN scheme is shown in simulation to outperform genetic-algorithm and random baselines, while also keeping higher-priority messages fresher than lower-priority ones.","feed_headline":"Deep RL cuts info age and energy in vehicle-to-vehicle links","feed_subtitle":"Multi-priority queues and NOMA let one roadside agent keep safety messages fresh while using less power than baselines.","key_machinery":"The central object is MPDQN, a deep Q-network variant for hybrid discrete-continuous action spaces: a policy network maps each discrete RRI choice to the best continuous transmit power, and a Q-network scores the resulting action tuples against the state. Two analytical models carry the argument: a multi-priority FIFO queue per vehicle that tracks AoI separately for each message type, and a NOMA/SIC receiver model that decodes the strongest colliding signal first and cancels it before decoding weaker ones, thereby raising SINR and reducing transmission failure. The collision probability itself is taken from a Markov-chain model of C-V2X Mode 4 resource selection, which sets how often the SIC decoder is exercised.","core_discovery":"The paper's central claim is that the resource reservation interval and transmit power of each vehicle can be selected online by a single DRL agent, trained with a multi-pass deep Q-network, to minimize the weighted sum of average receiver AoI and energy consumption in a C-V2X Mode 4 system. The agent observes per-vehicle state features, chooses among discrete RRI values (20, 50, or 100 ms) paired with a continuous power level, and receives a reward that is the negative weighted average of energy and AoI. Simulation results show the learned policy achieves lower average AoI and lower average energy consumption than genetic and random allocation, and that NOMA's SIC decoding reduces AoI most when the RRI is small and collisions are frequent.","pith_inferences":["The collision-probability model borrowed from [51] is assumed to remain valid when NOMA/SIC changes the interference structure; a more accurate SIC-aware collision model could shift the magnitude of the reported NOMA gains.","The fixed weighting factor ω1=0.6 implies a specific operator preference for energy savings over AoI; an adaptive weight that responds to the presence of high-priority traffic would be a natural extension of the reward design.","The discrete RRI set could be expanded or made continuous, which would let the agent trade off queue processing rate against collision frequency more finely than the current three choices allow.","The single-agent RSU architecture could be replaced by independent per-vehicle learners to test whether the performance degrades gracefully or requires coordination, a question the paper does not address."],"forward_implications":["A roadside unit could periodically recompute per-vehicle RRI and transmit power without centralized scheduling, adapting to vehicle density and channel conditions in real time.","Multi-priority queues guarantee that high-priority safety messages (HPD and DENM) maintain lower AoI than lower-priority CAM and MHD traffic, even when the total message arrival rate exceeds the processing rate.","NOMA with SIC can be layered on existing Mode 4 resource selection to reduce the damage of collisions, with the largest gains at short reservation intervals where collisions are most frequent.","Energy consumption stays nearly flat as the number of vehicles grows under the learned policy, while AoI increases only gradually, suggesting the scheme decouples energy cost from network load.","Because the policy is learned from local state at the RSU, the same architecture could be retrained for other road geometries, traffic patterns, or message-size distributions without changing the model."],"supporting_citations":[{"why":"Supplies the collision probability formula in Eq. (19), the key link between RRI choice, vehicle density, and the frequency of resource collisions.","marker":"[51]"},{"why":"Provides the LTE-V2V simulation platform that the paper modifies and extends for its own scenario, forming the basis of all simulation results.","marker":"[55]"},{"why":"Establishes the MPDQN method for jointly optimizing AoI and energy consumption in NR-V2X, which the paper adapts to the C-V2X Mode 4 setting.","marker":"[52]"},{"why":"Provides the justification and prior evidence that NOMA can alleviate latency and packet reception degradation in high-density V2X, supporting the paper's use of SIC-based NOMA.","marker":"[22]"},{"why":"Defines AoI as the timeliness metric and explains why lower delay and higher reliability both reduce average AoI, grounding the paper's performance objective.","marker":"[36]"},{"why":"Analyzes average AoI in multi-source queueing systems, which the paper builds on to construct its multi-priority queue AoI model.","marker":"[49]"}],"fun_headline_variants":["DRL tunes V2V power and scheduling to slash info age and energy","One DRL agent balances AoI and energy for C-V2X vehicles","NOMA plus DRL keeps V2V messages fresh while using less energy","Deep RL picks RRI and power to cut info age and energy in V2V","AI agent reduces age of info and power use in vehicle networks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the collision probability formula borrowed from [51] remains accurate when NOMA and SIC change how interfering signals are decoded; if that formula is wrong, the AoI improvements attributed to NOMA and the deep reinforcement learning policy are not established.","fun_headline_variants_meta":{"raw":{"variants":["DRL tunes V2V power and scheduling to slash info age and energy","One DRL agent balances AoI and energy for C-V2X vehicles","NOMA plus DRL keeps V2V messages fresh while using less energy","Deep RL picks RRI and power to cut info age and energy in V2V","AI agent reduces age of info and power use in vehicle networks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00024,"raw_usage":{"total_tokens":1548,"prompt_tokens":1007,"completion_tokens":541,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":441}},"tokens_in":623,"tokens_out":541,"duration_ms":5019,"temperature":1.0,"reasoning_tokens":441,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:49:33.358572+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a packet-level simulation of C-V2X Mode 4 with SIC-based NOMA and compare the measured fraction of collided resource reservations against the value predicted by Eq. (19) across the same vehicle densities and RRI values; a systematic discrepancy would invalidate the collision model on which the AoI and energy improvements rest.","supporting_citations":[{"cited_title":"Ltev2vsim: An lte-v2v simulator for the investigation of resource allocation for cooperative awareness,","cited_arxiv_id":null,"evidence_quote":"Provides the LTE-V2V simulation platform that the paper modifies and extends for its own scenario, forming the basis of all simulation results."},{"cited_title":"Joint optimization of age of information and energy consumption in nr-v2x system based on deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Establishes the MPDQN method for jointly optimizing AoI and energy consumption in NR-V2X, which the paper adapts to the C-V2X Mode 4 setting."},{"cited_title":"Non-orthogonal multiple access for high-reliable and low-latency v2x communications in 5g systems,","cited_arxiv_id":null,"evidence_quote":"Provides the justification and prior evidence that NOMA can alleviate latency and packet reception degradation in high-density V2X, supporting the paper's use of SIC-based NOMA."},{"cited_title":"Real-time status updating: Multiple sources,","cited_arxiv_id":null,"evidence_quote":"Analyzes average AoI in multi-source queueing systems, which the paper builds on to construct its multi-priority queue AoI model."}],"review_version":1}