{"id":"f4e96c4c-5aa2-48ef-a164-968a702cc3e4","arxiv_id":"2506.14237","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A new composite metric, Loss of Information Utility, combines delay and squared estimation error, and a semi-decentralized multi-agent reinforcement learning scheduler minimizes it for robot team D2D communications.","lead":"The paper introduces LoIU, a score that combines how old and how wrong a robot's knowledge of its teammates is, and uses that score to schedule robot-to-robot wireless transmissions under limited bandwidth. In simulations, the authors report large gains in information freshness and utility, and better task reliability than scheduling based on existing age-of-information metrics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (34) squares the posterior mean, which is zero under the paper's own Gaussian belief (Eq. 39), so it is not equivalent to the second-moment expression in Eq. (40); the expected-LoIU reward is therefore not derived as written.","rationale":"The reader identified the Wiener-process assumption and the zero-error reset as the weakest assumptions behind Eq. (40). Those are legitimate concerns about model realism and calibration. My concern is more immediate and internal: even granting every modeling assumption, Eq. (34) and Eq. (40) are not equivalent expressions. Eq. (34) is formed from the squared posterior mean, which is zero under the paper's own belief model, so the paper's derivation of the expected LoIU used in the MADDPG reward is mathematically invalid as written. This is load-bearing because the reward in Eq. (41) is exactly what the scheduler optimizes; if the derivation is not repaired, the link between the proposed metric and the learned policy is unsupported. The intended second-moment formula is plausible and likely what the implementation uses, which is why I would not move the verdict to REJECT. But the text must be corrected, and the authors should state explicitly whether Eq. (34) is a typo or whether the code uses the second moment directly. Because the reader's verdict was already CONDITIONAL, and my concern adds a sharper reason for that condition rather than changing the overall disposition, the appropriate verdict remains CONDITIONAL, i.e., UNCHANGED.","tokens_in":25826,"tokens_out":9041,"duration_ms":95425,"concrete_test":"Independently re-derive E[F_m(t)] from Eqs. (11) and (13) without using Eq. (34), and compare the result with Eq. (40). Then evaluate Eq. (34) for a single slot with xi_{n,m}=0 and the Gaussian belief of Eq. (39): Eq. (34) returns 0 while Eq. (40) returns (d_t_m/D_t_m) * (1/|C_m|) * sum_n (tau_{n,m}^t * sigma_n^2 / |E_{n,m}|^2). If the two results differ, the paper's derivation as written is wrong. As a second check, run the simulator with the reward replaced by Eq. (34) literally; if the learned schedules are indistinguishable from random or constant, the reported results cannot be attributed to the derivation in the paper until the second-moment form is justified directly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The closed-form expected LoIU in Eq. (40), and hence the MADDPG reward in Eq. (41), is presented as a rewriting of Eq. (34). It is not. Eq. (33) defines e_bar_{n,m} as the posterior mean E_{e~b}[e], and Eq. (34) squares that mean. Under the model's own assumption in Eqs. (9) and (39), the belief is Gaussian with zero mean whenever xi_{n,m}=0 and a point mass at zero when xi_{n,m}=1; hence e_bar_{n,m}=0 in both cases, and Eq. (34) is identically zero for every action. Eq. (40) instead uses E_{e~b}[|e|^2] = tau_{n,m}^t * sigma_n^2, which is the second moment of the belief and is nonzero for xi=0. The correct expected value of Eq. (13) is indeed the second-moment expression, but that expression must be derived from Eqs. (11)-(13), not from Eq. (34). As written, the derivation gives a degenerate reward: if Eq. (34) were what the scheduler optimized, the learning signal would carry no information about which transmissions to schedule. This is independent of the Wiener assumption; even if Eq. (9) is granted, the stated chain from Eq. (32) through Eq. (34) to Eq. (40) is invalid. The empirical 98% claim could survive if Eq. (40) is the implemented reward, but the paper's stated theoretical support for that reward is missing and must be corrected.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new metric, Loss of Information Utility (LoIU), which combines a normalized delay term and a squared normalized estimation error to quantify the freshness and accuracy of status information shared among collaborating robots. The authors formulate a scheduling and resource-allocation problem that minimizes time-average LoIU under latency constraints, reformulate it as a Dec-POMDP, and solve it with a semi-decentralized MADDPG algorithm in which each robot runs a local actor while a central critic is trained at the base station. The belief-based derivation of expected LoIU under a Wiener-process model leads to a closed-form reward, and simulations are used to compare the proposed scheme against DDPG, DQN, A2C, SAC, TDM, and threshold/random/all-allocated baselines, as well as against AoI, AoII, UoI, and AoCI as alternative metrics.","tokens_in":26176,"tokens_out":6676,"duration_ms":75975,"significance":"If the derivation and empirical claims are corrected, LoIU is a genuinely useful composite freshness-and-accuracy metric for robot-team cooperation, and the proposed semi-decentralized training framework addresses a real communication-overhead concern in multi-agent scheduling. The paper's strengths include an explicit system model, a closed-form expected-LoIU expression under a stated Wiener-process assumption, a clear problem formulation, a complexity analysis, and a broad set of comparisons against both prior metrics and prior RL baselines. The central theoretical derivation, however, currently contains a load-bearing error, and the convergence analysis is a sketch rather than a proof; these issues must be fixed before the results can be fully credited.","major_comments":[{"comment":"The derivation of the belief-based expected LoIU is invalid as written. Eq. (33) defines the average estimation error as the posterior mean e_bar = ∫ b(e) e de. Under the paper's own belief model, Eq. (38)-(39) give a zero-mean Gaussian when ξ=0 and a point mass at zero when ξ=1, so e_bar = 0 in every case. Consequently Eq. (34), which squares e_bar, is identically zero for every action and provides no learning signal. Eq. (40) instead uses E_b[|e|^2] = τ σ^2, which is the posterior second moment, not the square of the posterior mean. The correct expression can be obtained directly by taking the expectation of |e|^2 under the belief in Eq. (13); the chain through Eq. (34) must be rewritten, or the paper must clarify that Eq. (34) is not the expression used to define the reward. If Eq. (40) is what was implemented in the simulations, the stated theoretical support for that reward is currently missing.","section":"IV-B, Eqs. (32)-(40)"},{"comment":"The convergence proof does not establish convergence of the proposed semi-decentralized MADDPG. Definition 1 in Eq. (52) uses a strict inequality and describes a strict Nash equilibrium, not the Nash equilibrium needed for the subsequent argument. Because all agents share the same reward, Eq. (55) makes the potential function Φ in Eq. (54) coincide with the common reward, so the potential-game condition is satisfied trivially and only yields existence of an equilibrium in the finite stage game. The step from Eq. (56) to the almost-sure convergence statement in Eq. (57) assumes standard stochastic approximation conditions and Lipschitz/bounded-parameter conditions without verifying them for the specific intermittent-upload, semi-decentralized update scheme. The section should be reframed as a heuristic justification, or the missing conditions should be stated and verified.","section":"IV-D, Eqs. (50)-(57)"},{"comment":"The empirical validation is underreported. No error bars, confidence intervals, or numbers of independent seeds are given, and the abstract's claim of up to 98% improvement is not tied to a specific figure or table in Section V. Section V-D quotes one value (0.00506667) but does not show the 98% computation. Please report per-seed statistics, state the number of runs, and point to the exact comparison underlying the headline quantitative claims, or soften the claims accordingly.","section":"V, Figs. 3-5 and Tables IV-VI"}],"minor_comments":[{"comment":"The symbol F_m(t) is used for both the true LoIU and the belief-expected LoIU; please use separate notation such as E[F_m(t)] or a bar accent for the expected quantity.","section":"IV-B, Eqs. (34) and (40)"},{"comment":"In Definition 1, the condition should be rm(a*_m, a*_−m) ≥ rm(a_m, a*_−m) for all a_m, not the strict inequality written in Eq. (52), to match the verbal definition of Nash equilibrium.","section":"IV-D, Eq. (52)"},{"comment":"Table III lists 'penalty coefficient Rs 1', but the penalty coefficient in Eq. (24) is denoted ε; please align the notation and value.","section":"Table III"},{"comment":"The delay d_t_m is defined as the maximum over transmitted collaborators; if no collaborator is selected for transmission, d_t_m is undefined. Please state an assumption that at least one collaborator is selected or define d_t_m for the no-transmission case.","section":"III-B, Eq. (12)"},{"comment":"Two different metrics are both abbreviated AoCI: Age of Changed Information [31] and Age of Critical Information [32]. Please disambiguate these abbreviations in the table and text.","section":"Table II"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the proposed metric is potentially useful, but the derivation error in Section IV-B is central and must be corrected. The convergence analysis in Section IV-D also needs substantial revision. I see no reason to reject outright, provided the authors can fix the derivation and either strengthen the convergence argument or reframe it as a heuristic. The empirical claims should be backed by proper statistical reporting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a reasonable subfield contribution with a real hole in the derivation that needs to be fixed before publication.\n\nWhat is actually new: LoIU, a metric combining normalized delay with normalized squared estimation error under per-robot deadlines, plus a belief-distribution estimate of its expectation and a semi-decentralized MADDPG scheduler. That is a sensible extension of AoII and UoI with explicit collaboration semantics. The system model is clearly stated, and the comparison against AoI/AoII/UoI/AoCI is appropriate.\n\nThe load-bearing problem is in the expected-LoIU derivation. Eq. (34) evaluates the LoIU at the posterior mean \\bar{e}_{n,m} from Eq. (33). Under the paper's own Gaussian belief in Eq. (39), the posterior mean is zero whenever the update is missed, so Eq. (34) is identically zero for every action. Eq. (40) instead uses the second moment E[|e|^2], which is the correct expression for the expected squared error. The chain from Eqs. (32)-(34) to Eq. (40) is therefore invalid as written. The implemented reward is presumably the one in Eq. (40), which is right, but the paper does not derive it correctly. This is fixable with a clear derivation from Eq. (13), but it is a genuine error in the theoretical support.\n\nOther soft spots are less severe but worth noting. The convergence proof in Section IV-D is a sketch: it defines a potential function that is just the global reward, so it establishes existence of an NE for a one-step game, but it does not prove convergence of the MADDPG algorithm with experience replay and intermittent uploads. That is common in this literature, but the paper oversells it.\n\nThe headline numbers also do not line up. The abstract claims a 98% improvement in freshness and utility, while the task-reliability tables show gains of 3% to 12% over the second-best metric. The 98% may refer to LoIU values in Fig. 3, but the paper needs to reconcile the claims. No code, no error bars, and the Wiener-process assumption is stated but not tested against realistic mobility traces.\n\nCredit where due: the metric definition, problem formulation, and comparison setup are honest. The self-referential nature of evaluating a scheduler with its own metric is partially mitigated by the independent task-reliability measure.\n\nWho is this for: researchers working on AoI variants and multi-robot scheduling under wireless constraints. It deserves serious peer review because the core idea is useful and the main derivation error is repairable. I would send it out, but the authors must correct the derivation, add statistical rigor, and align the abstract with the reported numbers.","headline":"A plausible robot-team scheduling metric with a fixable but load-bearing derivation error: Eq. (34) squares the posterior mean, which is zero under the paper's own Gaussian belief, so the stated path to the reward in Eq. (40) is invalid.","tokens_in":26723,"tokens_out":2987,"would_cite":false,"duration_ms":40031,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper introduces LoIU, a metric multiplying update delay by normalized squared estimation error, and claims minimizing its time average with a semi-decentralized scheduler improves team information freshness and utility by up to 98%.","keywords":["Loss of Information Utility","robot teams","D2D communication","age of information","multi-agent deep reinforcement learning","transmission scheduling","belief distribution","resource allocation"],"falsifier":"Run the scheduler on a team whose statuses follow a different model—constant velocity with occasional sudden turns, or packet losses that leave residual error after a 'successful' reception—and compare the predicted expected LoIU from Eq. (40) with measured squared errors over many slots; a systematic mismatch that persists after retraining would show the central claim does not survive outside the Wiener-perfect-reception setting.","tokens_in":25638,"feed_emoji":"🤖","tokens_out":11558,"duration_ms":99633,"temperature":0.7,"pith_summary":"Robot teams depend on fresh and accurate status updates—position, velocity, intent—but wireless bandwidth is limited, so some updates wait and collaborators estimate instead. The paper proposes a single number, the Loss of Information Utility (LoIU), defined per robot as the product of a time-urgency factor (delay divided by its deadline) and a content-accuracy factor (the average squared estimation error normalized by its allowed bound). It argues that minimizing the team's time-average LoIU is the right objective for choosing which device-to-device transmissions to make and which resource blocks to use, because LoIU captures both how old and how wrong shared information is. The paper then shows how each robot can estimate its expected LoIU from a belief distribution of estimation errors, and designs a semi-decentralized multi-agent reinforcement learning scheduler to minimize it. If the claim holds, robot teams can keep shared knowledge fresh and accurate under bandwidth constraints, with task-reliability gains of roughly 33–44% over age-only or error-only metrics and up to 98% better information freshness and utility.","feed_headline":"New LoIU metric cuts robot-team information loss 98%","feed_subtitle":"One number weighs how old and how wrong shared robot status is, so a semi-decentralized scheduler can prioritize what matters.","key_machinery":"The load-bearing object is the LoIU functional $F_m(t)$ of Eq. (13): a deadline-normalized age factor multiplied by a threshold-normalized mean squared estimation error. Its companion is the belief-distribution estimator of Eq. (40), which turns the unobservable estimation error into a Gaussian with variance $\\tau_{n,m}^t \\sigma_n^2$ whenever an update is missed, making the expected LoIU computable from local observations. The third mechanism is the semi-decentralized MADDPG architecture: each robot runs an actor network for discrete collaborator and resource-block selection (made differentiable by Gumbel-Softmax), while a central critic at the base station trains on intermittently uploaded experiences and sends action gradients back to the actors, avoiding per-slot upload of full observations.","core_discovery":"The central claim is that one multiplicative indicator can rank the value of status updates for robot cooperation more usefully than age-only or error-only metrics. For robot $m$ at slot $t$, LoIU is $F_m(t) = (d_m^t / D_m^t) \\cdot (1/|C_m|) \\sum_{n \\in C_m} (e_{n,m}^t / E_{n,m})^2$, where $d_m^t$ is the delay of the update, $D_m^t$ its deadline, $e_{n,m}^t$ the estimation error of collaborator $n$'s status at robot $m$, and $E_{n,m}$ the maximum allowed error. Because robot statuses are modeled as independent Wiener processes, the expected value of this metric has a closed form: each unreceived update contributes $\\tau_{n,m}^t \\sigma_n^2 / E_{n,m}^2$, where $\\tau_{n,m}^t$ is the number of slots since the last successful reception and $\\sigma_n^2$ is the per-slot variance of the status process. The paper claims that using this expected LoIU as the reward in a semi-decentralized MADDPG scheduler—actors on robots choose collaborators and resource blocks, a central critic periodically refines them—reduces time-average LoIU and raises task reliability compared with AoI, AoII, UoI, AoCI, and single-agent reinforcement-learning baselines.","pith_inferences":["The LoIU definition itself does not depend on the Wiener assumption; replacing the Gaussian belief with jump-diffusion or learned predictive distributions and re-deriving the expected loss is a natural extension to robots with sudden maneuvers or discrete state changes.","Because LoIU separates into per-collaborator error terms multiplied by a common delay factor, it can scale to larger swarms by computing each link's contribution incrementally and aggregating locally.","The same indicator could serve as a reward-shaping term for other multi-agent coordination settings—aerial coverage, connected driving, remote surgery—whenever one agent's decision depends on estimates of other agents' states.","A direct calibration test would compare the predicted expected LoIU from Eq. (40) with realized squared errors in deployment; systematic underprediction after long gaps would signal that the constant-variance Wiener model needs per-robot adaptation."],"forward_implications":["A robot using LoIU will prioritize a collaborator whose status is both old and likely to be far off, rather than one that is merely old, so limited bandwidth goes to updates that most affect decision quality.","Expected LoIU is computable from local reception history and known status variance, so the same reward can be evaluated without waiting for ground-truth statuses, enabling decentralized execution.","The semi-decentralized design cuts per-slot upload and download traffic compared with centralized MADDPG, keeping the scheduler feasible when the air interface is congested.","Task reliability, defined as receiving or estimating collaborator status within both latency and error bounds, improves under LoIU relative to AoI-, AoII-, UoI-, and AoCI-based schedulers in the simulated AGV scenarios.","The closed form gives a direct, auditable relationship between missing an update and expected loss, so scheduling decisions can be explained rather than treated as a black box."],"supporting_citations":[{"why":"Defines Age of Information, the baseline freshness metric against which LoIU's task-reliability gains are measured.","marker":"[27]"},{"why":"Defines Age of Incorrect Information, a comparator metric that pairs age with an information penalty and is used in simulations.","marker":"[29]"},{"why":"Defines Urgency of Information, a context-aware comparator used to benchmark LoIU.","marker":"[30]"},{"why":"Defines Age of Changed Information, another content-aware comparator used in the task-reliability comparison.","marker":"[31]"},{"why":"Supplies the Wiener-process sampling model and the variance-growth law underlying the closed-form expected LoIU in Eq. (40).","marker":"[39]"},{"why":"Supports modeling multi-robot status fluctuations as Wiener processes, justifying the Gaussian belief assumption.","marker":"[42]"},{"why":"Provides the belief-distribution formalism used to estimate unobserved estimation errors when updates are missed.","marker":"[48]"},{"why":"The Gumbel-Softmax technique that lets the deterministic actor networks handle discrete collaborator and resource-block actions.","marker":"[49]"},{"why":"Establishes 0–1 integer programming as NP-complete, motivating the reduction of the scheduling problem to an NP-hard problem and the use of multi-agent reinforcement learning.","marker":"[45]"}],"fun_headline_variants":["New LoIU metric cuts robot-team info loss 98%","Robot team scheduler using LoIU boosts freshness 98%","LoIU: one number quantifies robot info utility, cuts loss 98%","Minimize robot info loss with new LoIU metric: 98% gain","Semi-decentralized LoIU scheduler improves robot info 98%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The scheme assumes a robot's status changes like a random walk with known, fixed step variance, and that a successfully received update is perfectly accurate; if statuses jump or updates arrive corrupted, the predicted information loss no longer matches reality.","fun_headline_variants_meta":{"raw":{"variants":["New LoIU metric cuts robot-team info loss 98%","Robot team scheduler using LoIU boosts freshness 98%","LoIU: one number quantifies robot info utility, cuts loss 98%","Minimize robot info loss with new LoIU metric: 98% gain","Semi-decentralized LoIU scheduler improves robot info 98%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000614,"raw_usage":{"total_tokens":2894,"prompt_tokens":1023,"completion_tokens":1871,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":639,"completion_tokens_details":{"reasoning_tokens":1791}},"tokens_in":639,"tokens_out":1871,"duration_ms":13354,"temperature":1.0,"reasoning_tokens":1791,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:17:58.875402+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the scheduler on a team whose statuses follow a different model—constant velocity with occasional sudden turns, or packet losses that leave residual error after a 'successful' reception—and compare the predicted expected LoIU from Eq. (40) with measured squared errors over many slots; a systematic mismatch that persists after retraining would show the central claim does not survive outside the Wiener-perfect-reception setting.","supporting_citations":[{"cited_title":"Real-time status: How often should one update?","cited_arxiv_id":null,"evidence_quote":"Defines Age of Information, the baseline freshness metric against which LoIU's task-reliability gains are measured."},{"cited_title":"The age of incorrect information: A new performance metric for status updates,","cited_arxiv_id":null,"evidence_quote":"Defines Age of Incorrect Information, a comparator metric that pairs age with an information penalty and is used in simulations."},{"cited_title":"Urgency of information for context- aware timely status updates in remote control systems,","cited_arxiv_id":null,"evidence_quote":"Defines Urgency of Information, a context-aware comparator used to benchmark LoIU."},{"cited_title":"Age of changed information: Content- aware status updating in the internet of things,","cited_arxiv_id":null,"evidence_quote":"Defines Age of Changed Information, another content-aware comparator used in the task-reliability comparison."},{"cited_title":"Sampling of the Wiener process for remote estimation over a channel with random delay,","cited_arxiv_id":null,"evidence_quote":"Supplies the Wiener-process sampling model and the variance-growth law underlying the closed-form expected LoIU in Eq. (40)."},{"cited_title":"Stochastic robust team tracking control of multi-UA V networked system under Wiener and Poisson random fluctuations,","cited_arxiv_id":null,"evidence_quote":"Supports modeling multi-robot status fluctuations as Wiener processes, justifying the Gaussian belief assumption."},{"cited_title":"Thrun, W","cited_arxiv_id":null,"evidence_quote":"Provides the belief-distribution formalism used to estimate unobserved estimation errors when updates are missed."},{"cited_title":"Categorical reparameterization with Gumbel-Softmax,","cited_arxiv_id":null,"evidence_quote":"The Gumbel-Softmax technique that lets the deterministic actor networks handle discrete collaborator and resource-block actions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes 0–1 integer programming as NP-complete, motivating the reduction of the scheduling problem to an NP-hard problem and the use of multi-agent reinforcement learning."}],"review_version":1}