{"id":"62057460-3e12-4bfc-8c72-4669f8f79692","arxiv_id":"2501.08020","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":9,"one_line_summary":"A VDPPO-based multi-agent model plans police patrol routes that cover most of the highest-crime nodes in two of three Malaga districts, with weaker results in the third.","lead":"This paper trains police patrol agents with multi-agent reinforcement learning to cover the highest-crime streets in three Malaga districts during a single shift. The best settings reach near-total coverage of the riskiest 3 percent of nodes in two districts, but the headline numbers do not hold in the third district.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The coverage index (Eqs. 7-9) is ambiguous: if G is the route-visited set the index is trivially 1; if G is the full candidate set, the stated definition lacks the normalization that produces the reported ratios. The abstract's 90%/65% claim rests entirely on this metric.","rationale":"The reader's weakest assumption is exactly the load-bearing problem: the formal definition of G in Equations 7-9 makes the coverage index either tautologically 1 or an unstated normalized ratio. I agree with that diagnosis and with the rejection. The concern is central because the abstract's two quantitative achievements (90% of the top 3% nodes, 65% of the top 20%) are expressed only in the units of this index. A corrected definition, plus a caption stating the denominator, might salvage the proposal; the repository and real-data setup are assets, and the greedy baseline is a reasonable sanity check. But as written, the primary evidence for the paper's core claim is internally inconsistent. The additional issues noted by the reader (zone 10 not reaching the claimed values, per-zone reward retuning, no error bars) are secondary to this metric ambiguity but reinforce the same conclusion: the paper currently overstates what is established. My recommended verdict is unchanged because the reader's REJECT is the appropriate classification for a manuscript whose headline result cannot be computed from its own equations.","tokens_in":23652,"tokens_out":5609,"duration_ms":56986,"concrete_test":"Run the public repository's evaluation and instrument the coverage-index computation for one configuration, e.g. zone 3, line of sight 1, best start, 5 patrols (Table 5, |W3|=0.999). Print the sets used: G, Z, Wψ, and the denominator. If G is the set of visited nodes, then Wψ must equal Z and the normalized value must be 1.000 for every row; if G is the full candidate set, the ratio |Wψ|/|Z| must equal 0.999. Whichever reading is implemented, check that the code's value matches the table under that reading; if neither definition reproduces the reported numbers, the metric is undefined and the headline claim cannot be checked.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5 defines the metric by saying 'Let G be the subset of nodes, G⊆C, to be monitored, covered in the routes of the agents' (Eqs. 7-8), then defines Wψ as the nodes in Z (the top ψ of G) visited in some episode (Eq. 9). If G is the union of nodes actually covered by the routes, every node of Z is covered by construction, so Wψ=Z and any normalized version of the index is 1 for every row. The tables report values from 0.141 to 1.000, which is impossible under that reading. The prose then calls the coverage index 'the cardinality of Wψ', but the table entries are fractional and cannot be cardinalities. The only consistent reading is that G is the full set of candidate nodes in the area and the reported number is |Wψ|/|Z|, but that denominator and interpretation are not stated. This is not a stylistic omission: the paper's headline numerical claims are quantified exclusively through this index, so an undefined or trivial metric means the central result cannot be verified from the text. Even under the charitable G=all-candidates reading, the abstract's unqualified 'more than 90% of the 3%' is contradicted by table 7, where the largest |W3| for zone 10 is 0.712, so the claim is at best restricted to zones 3 and 9.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses the problem of planning coordinated patrol routes in urban environments using multi-agent reinforcement learning. The authors model the problem as a decentralized partially observable Markov decision process (dec-POMDP) on an undirected graph and train a shared policy with the VDPPO algorithm. The approach is evaluated on three districts of Málaga using historical crime data, and a new 'coverage index' inspired by the predictive accuracy index (PAI) is introduced to measure how well the generated routes visit the highest-crime nodes. The authors also compare against a greedy baseline and analyze the effects of the number of patrols, line of sight, and initial deployment strategy. The abstract claims that the coordinated routes achieve more than 90% coverage of the top 3% highest-crime nodes and 65% coverage of the top 20%, for the tested scenarios.","tokens_in":24001,"tokens_out":9416,"duration_ms":87319,"significance":"If the results are validated, the paper would be a relevant contribution to MARL-based patrol planning, offering a real-world testbed, a reproducible codebase (public repository), and a practical resource-allocation question about the required number of patrols. The motivation for a coverage metric suited to finite-horizon, non-cyclical routes is reasonable. However, the current significance is limited by the fact that the coverage index is not consistently defined in the text, the headline claim is contradicted by the results for zone 10, the reported point estimates lack uncertainty quantification, and the evaluation metric and training reward share the same target function, making the results largely a measure of optimization success rather than an independent assessment. With substantial revision, the approach could be of interest to both the applied reinforcement learning and crime-prevention communities.","major_comments":[{"comment":"The definition of the coverage index is ambiguous and inconsistent with the reported tables. The text defines G as 'the subset of nodes ... covered in the routes of the agents'; if this is interpreted literally, every node of Z (a subset of G) is visited by construction, so Wψ = Z and the index is trivially 1 after any reasonable normalization. The tables report fractional values between 0.141 and 1.000, which cannot be cardinalities of a subset of nodes. The only reading consistent with the numbers is that G is the full set of candidate nodes in each zone and the reported value is |Wψ| / (ψ|G|/100) (or an equivalent proportion of the top-ψ nodes that are visited), but this denominator and normalization are not stated anywhere. Because every headline claim is expressed through this index, the paper must define it unambiguously and consistently with the tables.","section":"Section 5, Eqs. (7)-(9)"},{"comment":"The abstract's claim of 'more than 90% of the 3% of graph nodes with the highest crime incidence' is not supported by the results for zone 10, where the largest |W3| value is 0.712 (Table 7, line of sight 6, random initial positions, 5 patrols). Even under the charitable reading of the coverage index, this claim holds only for zones 3 and 9, and only for certain configurations. The abstract should be revised to state the claim as zone-specific (e.g., 'for zones 3 and 9') or to describe the range of results, such as 'up to more than 90%'.","section":"Abstract vs. Table 7"},{"comment":"The tables report a single point estimate per configuration with no standard deviation, confidence interval, or statistical test, despite the text mentioning '100 runs' in Section 5. Without measures of variability, the reader cannot assess whether the observed differences across line-of-sight values, numbers of patrols, and initial positions are meaningful, nor whether the model's improvements over the greedy baseline are significant. The paper should provide error bars or the full distribution of outcomes, and where relevant, apply significance tests or report effect sizes.","section":"Tables 5-7"},{"comment":"The evaluation metric and the training reward both rely on the same target function σ (crime incidence): the reward function (Eq. 5) directly rewards visiting nodes with high σ, and the coverage index (Eqs. 7-9) measures the fraction of high-σ nodes visited. In addition, the 'best' initial placement selects nodes with the highest σ. Consequently, the reported coverage largely reflects the agents' ability to optimize the training objective rather than an independent measure of surveillance effectiveness. The greedy baseline is a useful external reference, but it is also defined in terms of σ. To support the broader claim of 'optimizing urban crime surveillance,' the paper should provide out-of-sample or temporal validation (e.g., training on earlier years and evaluating on later crime data) or explicitly limit the claim to coverage of known historical hotspots.","section":"Section 3.5 and Section 5"},{"comment":"The paper reports that only 304,125 of 376,737 crimes (about 19%) were successfully geolocated and used for the target function σ, but it does not discuss whether the omitted crimes are spatially biased. If the geocoding failures are not spatially random, σ will be a distorted representation of crime density, and all coverage results will inherit that distortion. The paper should analyze geocoding completeness by zone, discuss the potential for spatial bias, or apply an appropriate correction, and in any case should state this as a limitation explicitly.","section":"Section 4.1"}],"minor_comments":[{"comment":"The Conclusions section contains several typographical errors and garbled passages, such as 'we counted with rreliedaonthe crimes comed in the city' and 'WFurthermore,e have also introduced'; this section needs a careful copyedit.","section":"Section 7"},{"comment":"The subsection heading 'umber of patrols' in the Discussion is missing the initial 'N'; it should read 'Number of patrols'.","section":"Section 6"},{"comment":"Equation (1) contains an apparent typo in the recursive case: the condition 'vk∈V,vk,vi' should likely be 'vk ≠ vi' to correctly express the shortest-path recurrence.","section":"Eq. (1)"},{"comment":"The notation |Wψ| implies a cardinality, but the reported values are normalized fractions. Once the coverage index is formally redefined, the authors should use a distinct notation (e.g., Cψ or a fractional coverage) to avoid confusion.","section":"Section 5"},{"comment":"The paper does not specify whether the reported coverage values are means, medians, or maxima over the 100 runs; please clarify this for both the coverage and entropy columns.","section":"Tables 5-7"},{"comment":"The concept of 'line of sight' is not formally defined; the paper should specify the exact geometry (e.g., a square of side 2k+1 cells centered on the agent) and whether the range is measured in grid cells.","section":"Section 3.5 and 5"},{"comment":"The column 'Preselected or homogeneous nodes' is ambiguous; consider splitting it into 'Preselected nodes' and 'Homogeneous node weights' for clarity.","section":"Table 1"},{"comment":"The caption says 'Three of the five routes' but does not identify the configuration (zone, number of patrols, line of sight); please specify these details.","section":"Figure 9"},{"comment":"The paper states that for zone 10 'the results with 10 patrols surpass those with 2 and 5 patrols' after changing the reward parameters, which makes cross-zone comparisons of the 'optimal number of patrols' conclusion difficult; this sensitivity should be acknowledged in the Discussion.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of an applied AI/RL venue and the authors provide a useful real-world testbed. However, the central evaluation metric is not formally well-defined as written, and the abstract overstates the results relative to the tables. I do not recommend rejection because the definitional issue appears fixable and the experimental setup is otherwise reasonable, but the revised version must present an unambiguous coverage index, include uncertainty quantification, and qualify the claims appropriately. I would also encourage the editor to request that the authors either release an anonymized version of the data or provide a detailed analysis of the geocoding bias, given the confidentiality statement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: this is a genuine applied-MARL paper with real data from three Malaga districts, code released, and an honest comparison against a greedy baseline. The authors know the patrol-routing literature and make a fair case that idleness metrics don't fit single-shift, no-preselection patrol. The VDPPO vs IPPO vs MAPPO comparison is useful. If you work on police routing or hotspot coverage, it's worth a skim.\n\nThe problem is the coverage index. Section 5 defines G as \"the subset of nodes ... covered in the routes of the agents,\" then defines Z as the top ψ% of G. If that's what G is, every node in Z is visited by construction and the index is trivially 1. The tables show values like 0.14 to 1.00, so that reading can't be right. The only consistent reading is that G is the full set of candidate nodes and the reported number is |Wψ|/|Z|, but that denominator is never stated. The text also calls the index \"the cardinality of Wψ,\" which conflicts with the fractional table entries. That's not a stylistic quibble: the whole headline result is expressed through this index.\n\nThe abstract says \"more than 90% of the 3% ... and 65% for 20%.\" Table 7 (zone 10) tops out at 0.712 for |W3| and 0.538 for |W20|. So the claim only holds, at best, for zones 3 and 9 with favorable configurations. The authors didn't qualify it.\n\nOther soft spots: reward parameters were retuned per zone (zone 10 got a different coverage factor and doubled exploration rewards), which weakens the claim that the model generalizes. Tables report point estimates with no error bars or significance tests. The crime data were geolocated only for 304,125 out of 376,737 reports, a 19% exclusion that could bias the target function. Data are withheld for confidentiality, but code is public.\n\nNone of this is fatal to the underlying idea. The metric is fixable by stating G correctly and reporting |Wψ|/|Z| with error bars; the abstract needs zone-qualified claims; and a temporal out-of-sample test would strengthen the \"coverage\" story. But as written, the central quantitative claims cannot be verified from the text.\n\nRecommendation: send it to peer review with a request for major revision. A good referee can push for the corrected metric, qualified abstract, and proper uncertainty quantification. It's not a desk reject — there's a real application and honest work underneath — but it's not publishable in this form.","headline":"Real applied-MARL work with a useful greedy baseline, but the coverage index is ill-defined and the abstract's 90%/65% numbers are contradicted by Table 7.","tokens_in":24526,"tokens_out":3350,"would_cite":false,"duration_ms":30803,"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":"Trained on real crime data from three Malaga districts, a cooperative multi-agent reinforcement learning model plans single-shift patrol routes that cover over 90% of the highest-crime cells.","keywords":["multi-agent reinforcement learning","patrol routing","coverage index","crime hotspots","urban surveillance","dec-POMDP","VDPPO","cooperative routing"],"falsifier":"Recompute the coverage index for any reported configuration using the full set of nodes in the zone as the denominator; if the fraction of top-3% nodes visited does not reproduce the table values (for example, 0.999 for zone 3 with five patrols, line of sight 3, best starting positions), then the formal definition of G as the route-visited subset is the one being used and the headline percentages describe something weaker than district-wide coverage.","tokens_in":23415,"feed_emoji":"🚔","tokens_out":6465,"duration_ms":58325,"temperature":0.7,"pith_summary":"The paper claims that a cooperative multi-agent reinforcement learning model, trained with VDPPO, can turn real crime data into coordinated patrol routes for a single eight-hour shift. In three medium-sized districts of Malaga, the resulting routes cover more than 90% of the top 3% of cells by crime incidence and about 65% of the top 20%. To make this measurable, the authors introduce a coverage index inspired by criminology's predictive accuracy index. They also show that five patrols perform almost as well as ten, and that starting patrols at known hotspots beats random deployment.","feed_headline":"MARL patrol routes cover 90% of worst crime hotspots","feed_subtitle":"VDPPO-trained agents plan coordinated single-shift routes for Malaga districts; five patrols match ten.","key_machinery":"The machinery is VDPPO (Value Decomposition Proximal Policy Optimization), a multi-agent PPO variant that decomposes the joint value function to handle credit assignment among homogeneous agents. Observations are filtered through an agent line of sight, and the reward function $R'_t(a_i)$ combines a node-value term with an exploration term $\\tau_t$ and a coverage penalty, summed with the joint reward to discourage free-riding. The paper's evaluation metric is the coverage index $|W_\\psi|$, which counts how many of the $\\psi$% highest-crime nodes in the area are visited by any patrol; $\\psi$ is typically 3% to 20%.","core_discovery":"The central discovery is that route design for urban police patrols can be posed as a decentralized partially observable Markov decision process and solved with a cooperative multi-agent reinforcement learning algorithm, Value Decomposition PPO, without preselecting which nodes to visit. Agents share a policy, observe patrol positions, recent visit counts, and target values within a limited line of sight, and are rewarded through a combination of individual and joint coverage terms designed to prevent lazy agents and encourage exploration. Evaluated on 50-meter grid graphs built from real street maps and 304,125 geolocated crimes in Malaga, the trained policies generate routes that concentrate on high-crime cells. The authors report that the best configurations cover essentially all of the top 3% highest-crime nodes in the denser districts and still cover roughly 65% of the top 20% in the larger, sparser district.","pith_inferences":["An implication left implicit is that the model could serve as a pre-shift planning tool rather than real-time dispatch, since the paper assumes patrols will not deviate from their route for incidents.","The same coordinated-route machinery should transfer to other spatially heterogeneous surveillance tasks, such as drone inspection or environmental monitoring, but the reward parameters would need to be re-tuned for each new target function, as the paper itself notes.","A concrete extension would be to couple the trained policy with a fast local replanner that handles emergency deviations, which the authors list as future work and which the current model explicitly does not support."],"forward_implications":["Cities can estimate the right number of patrols: in the two denser districts, five patrols achieve coverage nearly equal to ten, so the extra five mostly add cost without adding surveillance.","Agents do not need full information: a line of sight of 3 performs about as well as 6, which supports deployment with limited reconnaissance.","Randomizing starting positions buys little: it raises route entropy by less than 10% and generally lowers coverage compared with starting patrols at the highest-crime nodes.","The learned routes beat the greedy nearest-high-value-cell baseline in almost every configuration, and clearly so in the two smaller districts.","Sparse, large districts are harder: zone 10 needed stronger exploration rewards and still achieved lower coverage, indicating that per-area parameter tuning is required when the model is transferred."],"supporting_citations":[{"why":"Supplies the dec-POMDP formalism used to define the patrol problem.","marker":"Oliehoek et al., 2016"},{"why":"Provides the predictive accuracy index (PAI) that the coverage index is based on.","marker":"Chainey et al., 2008"},{"why":"Introduces VDPPO, the algorithm that performed best in the experiments.","marker":"Ma and Luo, 2022"},{"why":"Supplies the underlying PPO method that VDPPO and the independent baselines extend.","marker":"Schulman et al., 2017"},{"why":"Provides the MARLlib implementation used to train the compared algorithms.","marker":"Hu et al., 2023"},{"why":"Defines the related MARL patrolling baseline balancing efficiency and unpredictability that motivates the design.","marker":"Guo et al., 2023"}],"fun_headline_variants":["MARL patrols cover 90% of worst crime nodes","AI cops hit 90% of worst crime hotspots","Coordinated patrol routes nail 90% of top crime cells","Reinforcement learning steers patrols to 90% crime coverage"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline coverage numbers assume that the metric's denominator is the full set of monitorable cells in the district, whereas the paper's formal definition says the denominator is only the cells the patrols actually visited, which would make the index trivially 1.","fun_headline_variants_meta":{"raw":{"variants":["MARL patrols cover 90% of worst crime nodes","AI cops hit 90% of worst crime hotspots","Coordinated patrol routes nail 90% of top crime cells","Reinforcement learning steers patrols to 90% crime coverage"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000904,"raw_usage":{"total_tokens":3945,"prompt_tokens":1059,"completion_tokens":2886,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":675,"completion_tokens_details":{"reasoning_tokens":2815}},"tokens_in":675,"tokens_out":2886,"duration_ms":19454,"temperature":1.0,"reasoning_tokens":2815,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:30:35.646022+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the coverage index for any reported configuration using the full set of nodes in the zone as the denominator; if the fraction of top-3% nodes visited does not reproduce the table values (for example, 0.999 for zone 3 with five patrols, line of sight 3, best starting positions), then the formal definition of G as the route-visited subset is the one being used and the headline percentages describe something weaker than district-wide coverage.","supporting_citations":[],"review_version":1}