{"id":"2bafc854-1cf1-4d85-9c92-fa7e61b9c6e2","arxiv_id":"2607.22667","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A PPO policy selecting one sensor per second from particle-filter belief features is statistically equivalent to explicit information-gain selection and non-inferior to all-sensors-on within 2 m RMSE / 2% lost-track margins in a simulated marina.","lead":"This paper trains a reinforcement-learning agent (PPO) to pick one sensor at a time — out of three cameras and two LiDARs — for tracking a small boat in a simulated marina, instead of running all sensors at once. In the simulation the one-sensor policy tracks nearly as well as using all five sensors, while costing far less computation than the explicit information-gain baseline.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'close to always-on' result is untested outside the paper's limited-overlap sensor layout; the central claim is conditional on that geometry, as the authors acknowledge.","rationale":"The reader's weakest-assumption section identifies both sensor-geometry dependence and simulation fidelity. I find the geometry dependence more load-bearing for the central claim as stated: the paper's own mechanism predicts the near-equivalence will weaken as FOV overlap increases, but the paper does not test that prediction. Simulation fidelity is a real external-validity limitation, but the paper consistently frames the study as a georeferenced simulation and lists real-world validation as future work; that is a disclosed scope boundary rather than an untested assumption that the central claim silently depends on. The 2 m/2% margins without operational derivation are also a concern, but they primarily affect how 'close' is interpreted; the geometry issue affects whether 'close' survives the paper's own proposed next step. I therefore maintain the reader's CONDITIONAL verdict: the conditions already listed (alternative geometry, multi-seed, code/data, real sensor-model validation) are exactly what would resolve this concern. No change to the verdict is needed.","tokens_in":9076,"tokens_out":5389,"duration_ms":58135,"concrete_test":"Run the identical IG-RLSS training and evaluation pipeline on a modified CMMI Smart Marina layout with deliberately high FOV overlap — e.g., add two cameras/LiDARs so that every water position within the operational region is covered by at least two sensors simultaneously. Keep the same particle filter, reward, PPO hyperparameters, and 1,000 held-out routes. If IG-RLSS remains non-inferior to always-on within the same 2 m / 2% margins in all zones, the geometry-dependence concern is refuted. If the gap exceeds those margins in any zone, the paper's conclusion should be re-scoped to limited-overlap sensor configurations, and the abstract/summary should say so.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that IG-RLSS 'achieves tracking performance close to always-on sensing' while activating one sensor, operationalized as TOST/non-inferiority within 2 m RMSE and 2% lost-track. The paper's own mechanistic explanation is that this holds because the fixed sensors' FOVs overlap only over a limited portion of the operational area; for most target locations the PF update is dominated by a single observing sensor or sensors providing 'limited additional independent information' (Section III, close-performance discussion). With a denser, redundant layout, always-on and EIG-SS would supply genuinely independent measurements, and the one-sensor policy's equivalence could exceed the pre-specified margins. The authors list exactly this as future work ('alternative static sensor configurations ... ranging from sparse complementary coverage to highly redundant coverage'), effectively acknowledging that the demonstrated closeness is tied to the tested geometry. Since no alternative geometry is evaluated, the headline claim is conditional on that geometry. This is a scoping gap, not internal inconsistency.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes IG-RLSS, a reinforcement-learning framework for selecting one of five fixed camera/LiDAR sensors at each decision step during single-vessel tracking in a simulated georeferenced marina. A particle filter maintains the belief state, and a PPO agent is trained with a reward equal to realized posterior information gain gated by an observability mask. The method is compared against random single-sensor selection, always-on sensing, and an expected-information-gain baseline (EIG-SS) on 1,000 held-out test routes. The authors report that IG-RLSS is TOST-equivalent to EIG-SS within 2 m RMSE and 2% lost-track margins in all three spatial zones, non-inferior to always-on at the same margins, and cheaper in FLOPs than EIG-SS. They also explain that the close performance is tied to the limited overlap of the sensor layout and propose more redundant configurations as future work.","tokens_in":9243,"tokens_out":4830,"duration_ms":50377,"significance":"If the results are reproduced, the paper offers a practical amortized alternative to online expected-information-gain sensor scheduling in resource-constrained maritime surveillance, with a clear computational advantage. The study is carefully designed in several respects: matched-pair statistics on pre-generated test routes, pre-specified equivalence and non-inferiority margins, held-out evaluation, and an explicit FLOP comparison. The authors also honestly flag the dependence of the result on the specific sensor geometry and the absence of real-world validation. These strengths make the contribution worthwhile, provided that the definitional and reporting gaps identified below are addressed.","major_comments":[{"comment":"The realized information gain I_real^t, the EIG-SS baseline, and the lost-track percentage are all defined by reference to the authors' prior work [5] rather than in this manuscript. These quantities are load-bearing: the reward drives training, EIG-SS is the central comparison, and lost-track is a headline metric. Without explicit definitions the method is not reproducible and the reader cannot assess whether the reported equivalence is meaningful. The manuscript should include self-contained definitions of I_real^t, the EIG-SS selection rule, and the lost-track metric.","section":"Section II, reward paragraph; Section III, evaluation metrics"},{"comment":"The paper states that TOST equivalence and one-sided non-inferiority tests show IG-RLSS remains within the 2 m / 2% margins in all zones, but it does not report the test statistics, confidence intervals, or p-values. The reader cannot verify the central claim. Provide a table with, for each zone and comparison, the estimated paired difference, the 90% (or 95%) confidence interval, the pre-specified margin, and the TOST/non-inferiority conclusion.","section":"Section III, statistical tests paragraph"},{"comment":"The abstract and conclusion claim that the learned policy achieves tracking performance close to always-on sensing as a general result, but the paper's own mechanistic explanation is that this holds because the fixed sensors' fields of view overlap only over a limited portion of the operational area; for most target locations a single sensor dominates the update. The only evaluated layout is the low-overlap marina sensor configuration. The claim should be explicitly scoped to this geometry, or a complementary experiment with a redundant, high-overlap layout should be added. This is a scoping gap, not an internal inconsistency, but it is central to how the result is presented.","section":"Abstract and Conclusion vs. Section III close-performance discussion"}],"minor_comments":[{"comment":"Typo: 'rather than from an ground-truth tracking-error penalty' should be 'rather than from a ground-truth tracking-error penalty'.","section":"Section IV, Conclusion"},{"comment":"Spelling of LiDAR is inconsistent (both 'LiDAR' and 'LIDAR' appear). Use one standard form.","section":"Throughout"},{"comment":"The phrase 'one-sided non-inferiority tests (TOST)' is imprecise: TOST is a two one-sided tests procedure for equivalence, whereas non-inferiority is a one-sided test. Clarify the terminology, e.g. 'one-sided non-inferiority tests' without labeling them TOST.","section":"Section III, statistical tests paragraph"},{"comment":"The FLOP comparison is useful but the paper acknowledges it excludes runtime, memory, and energy. Consider adding a sentence in the caption or text noting that these are implementation-dependent; the current text already does this, so this is only a presentational suggestion.","section":"Section III, Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"The core idea and evaluation protocol are sound, and the scoping limitation is acknowledged by the authors. The major issues are definitional incompleteness (relying on an unpublished overlapping-author preprint for the reward, baseline, and metrics) and insufficient reporting of the statistical evidence for the central equivalence/non-inferiority claims. Both are fixable within the manuscript's scope. I do not see a fatal flaw, but the revision must include full definitions and a proper table of test statistics before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, honest engineering paper, not a paradigm shift. The new thing is the specific integration — a PPO policy selecting one fixed sensor per second from a particle-filter belief, trained with a realized-IG reward, and compared against an explicit EIG-SS baseline. Prior RL scheduling work mostly used non-SMC beliefs or mobile-agent control, so this is a legitimate new combination. The paper does not oversell it, which I appreciate.\n\nWhat it does well: the statistical work is unusually careful for this area. Matched-pair tests, Holm correction, TOST equivalence to EIG-SS and non-inferiority to always-on at pre-specified 2 m / 2% margins on 1,000 held-out routes. They also disclose the Zone 3 degradation honestly — all methods lose track about 45% of the time there, and they show it is a coverage problem, not a scheduling problem. The mechanistic explanation for why one-sensor selection nearly matches always-on — limited FOV overlap in this particular marina layout — is clear and plausible. The FLOP estimate is rough, but they label it as arithmetic cost only.\n\nThe soft spots, in proportion. Biggest scoping gap: the \"close to always-on\" claim is conditional on the sensor geometry. The authors say so themselves, and they list alternative geometries as future work, but they never test one. With a denser, redundant layout the equivalence could collapse. That is not a fatal flaw — it is a scoping limitation they explicitly acknowledge. Second, all results are simulation-only, with no code or data release and no multi-seed training variance reported. That limits reproducibility and makes it hard to judge how robust the policy is. Third, the equivalence margins (2 m, 2%) are asserted as application-level tolerances without derivation from operational requirements. Minor: the reward is defined the same way as in their prior EIG-SS work [5], which is an overlapping-author citation, but that is by design — the learned policy is meant to amortize that very criterion.\n\nWho this is for: anyone working on practical sensor management, especially maritime surveillance with fixed camera/LiDAR networks, and anyone who wants a clean example of combining SMC belief representations with RL. It deserves a serious referee. My recommendation: send it to peer review, but require the authors to release the simulator and evaluation code, report multiple training seeds, justify the margins, and ideally evaluate at least one alternative sensor geometry — even a stylized one. Those are reasonable conditions, not showstoppers.","headline":"A careful, honest RL sensor-scheduling paper whose central result is real but explicitly tied to its limited-overlap sensor geometry; deserves peer review with requests for code, seeds, and a second geometry.","tokens_in":9868,"tokens_out":1673,"would_cite":true,"duration_ms":19067,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A trained policy that activates one sensor per step can track vessels almost as well as always-on sensing, while avoiding the online entropy search of expected-information-gain selection.","keywords":["sensor scheduling","reinforcement learning","particle filter","information gain","maritime surveillance","target tracking","heterogeneous sensor network","proximal policy optimization"],"falsifier":"Run the identical IG-RLSS training and test procedure on a simulated sensor layout with heavily overlapping fields of view (for example, co-located cameras and LiDARs covering the same area) and check the TOST margins: if IG-RLSS is no longer within 2 m RMSE / 2% lost-track of always-on sensing in the covered regions, the claimed near-equivalence is specific to sparse coverage and does not generalize.","tokens_in":8869,"feed_emoji":"📡","tokens_out":7370,"duration_ms":67105,"temperature":0.7,"pith_summary":"This paper tries to establish that a reinforcement-learning policy can decide which single sensor to activate at each decision step in a fixed heterogeneous maritime sensor network and still track a vessel nearly as well as activating all sensors at once. The proposed framework, IG-RLSS, trains a proximal-policy-optimization agent on particle-filter belief features, rewarding the realized reduction in posterior entropy. On 1,000 held-out test routes in a simulated marina, the learned single-sensor policy is statistically equivalent to expected-information-gain selection within 2 m RMSE and 2% lost-track margins in all zones, and non-inferior to always-on sensing at the same margins, while costing fewer FLOPs than the online entropy-search baseline. The practical point is that a fixed camera-LiDAR network can run at a fraction of the sensing and computation cost without sacrificing tracking quality, at least when fields of view overlap only sparsely.","feed_headline":"One sensor per step stays within 2 m of full-sensor tracking","feed_subtitle":"Learned sensor policy skips online entropy search and stays within 2 m of always-on tracking.","key_machinery":"The load-bearing mechanism is an amortized information-gain scheduler: a PPO policy that maps a belief-derived observation vector to a one-of-five sensor action, trained with a reward equal to the realized posterior entropy reduction of the particle filter, multiplied by an observability mask indicating whether the vessel is seen by at least one sensor. The particle filter (500 particles, kinematic motion with a turn-rate constraint, Gaussian Cartesian likelihoods, missed-detection updates) supplies the belief and the information-gain signal, while the policy supplies the selection decision. This replaces the explicit counterfactual entropy search of EIG-SS with a single forward pass of a tw","core_discovery":"On the paper's own terms, the central discovery is that explicit online information-gain search is not necessary to obtain near-optimal sensor scheduling: a policy trained offline with a realized-information-gain reward gated by an observability mask can amortize the selection decision. Using a 500-particle sequential Monte Carlo tracker as the belief source and a 56-dimensional observation vector encoding belief spread, detection history, coverage, and sensor geometry, the PPO agent selects one of five sensors each second. Over the held-out test split, IG-RLSS is TOST-equivalent to EIG-SS within 2 m RMSE / 2% lost-track in all zones, non-inferior to always-on sensing at the same margins, an","pith_inferences":["If this carries over to other sparse-overlap networks, it implies that the expensive part of information-driven sensing is the online search, not the information criterion; the same realized-IG reward could be used to train subset-selection policies directly.","A natural stress test is a high-overlap layout: the paper's own geometry explanation predicts that the equivalence to always-on should break down as redundant sensors begin contributing independent information; running the same training procedure on such a layout would either confirm the boundary or falsify the generalization.","The reward is posterior-entropy reduction, not tracking error, so the policy may be choosing sensors that reduce uncertainty even when that does not reduce RMSE; comparing against an error-reward variant could separate informative from accurate behavior.","Because all results are simulation-based, the strongest testable extension is deployment on a real camera-LiDAR marina testbed; unmodeled effects such as occlusion, weather, and calibration drift could change which sensor is genuinely informative."],"forward_implications":["In networks with sparse field-of-view overlap, single-sensor scheduling can recover nearly all of the tracking benefit of always-on sensing; the residual difference is within 2 m RMSE and 2% lost track.","A learned policy can serve as a low-compute substitute for expected-information-gain selection, avoiding repeated counterfactual particle-filter updates at run time.","Tracking quality in the far field is limited mainly by coverage rather than by the scheduling rule: all non-random methods degrade similarly, and no sensor covers the true position during about 18% of far-zone timesteps.","The framework establishes a reusable belief-based, single-sensor scheduling baseline and simulation environment for extending to multi-sensor subset selection and energy-aware policies."],"fun_headline_variants":["One sensor matches full array within 2 m","Learned policy picks one sensor, skips costly search","Single-sensor RL tracks within 2 m of always-on","Offline RL sensor choice beats online entropy search","Near-perfect tracking with one sensor and no search"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central claim rests on the sensor layout being one where a single sensor dominates the particle-filter update for most target locations; in a denser, higher-overlap layout the equivalence to always-on sensing could break, and the simulator's fidelity to a real marina remains unvalidated.","fun_headline_variants_meta":{"raw":{"variants":["One sensor matches full array within 2 m","Learned policy picks one sensor, skips costly search","Single-sensor RL tracks within 2 m of always-on","Offline RL sensor choice beats online entropy search","Near-perfect tracking with one sensor and no search"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000163,"raw_usage":{"total_tokens":1090,"prompt_tokens":765,"completion_tokens":325,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":509,"completion_tokens_details":{"reasoning_tokens":262}},"tokens_in":509,"tokens_out":325,"duration_ms":4189,"temperature":1.0,"reasoning_tokens":262,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T08:52:14.567540+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical IG-RLSS training and test procedure on a simulated sensor layout with heavily overlapping fields of view (for example, co-located cameras and LiDARs covering the same area) and check the TOST margins: if IG-RLSS is no longer within 2 m RMSE / 2% lost-track of always-on sensing in the covered regions, the claimed near-equivalence is specific to sparse coverage and does not generalize.","supporting_citations":[],"review_version":1}