{"id":"178690e7-fcad-4c0e-9789-881b9336adf9","arxiv_id":"2506.03663","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"An improved grey wolf optimizer with a centroid-based cooperative update and lens opposition-based learning gives shorter planned UAV paths than GWO, PSO, and WOA in four simulated maps.","lead":"This paper proposes an improved grey wolf optimizer that adds a group-centroid update and an opposition-based learning step, then applies it to UAV shortest path planning. The authors report that it beats three standard metaheuristics on several benchmark functions and produces slightly shorter paths in four simulated obstacle maps.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The path-planning objective in Eq. (11) penalizes obstacle hits with P×n_O and P=10, yet feasible path lengths in the 20m x 20m maps are about 27–31m; a path intersecting one obstacle cell scores 10 and beats every feasible path, so Table III may not describe collision-free paths.","rationale":"The reader's weakest assumption correctly notes the absence of repetitions, standard deviations, and statistical tests in the path-planning comparison, and that is a valid concern. However, the more fundamental defect is in the objective function itself. Section II-B defines f(X)=P×n_O whenever n_O>0, with P=10 in Section III-B. On the 20m×20m maps, feasible collision-free paths have lengths near 27–31m. Therefore a path that intersects a single obstacle cell receives objective value 10, which is less than every feasible path's objective. Minimizing Eq. (11) will favor infeasible shortcuts, so the path lengths reported in Table III may not correspond to obstacle-free routes. This is an internal inconsistency, not a matter of external consensus or baseline tuning. Even if all four algorithms were rerun 30 times with tuned baselines, the flaw would remain unless the penalty is corrected and feasibility is explicitly verified. The paper's central applied claim, that IGWO plans shorter UAV paths, is therefore not supported by the experiments as described. I recommend REJECT rather than CONDITIONAL because the current objective function actively rewards obstacle intersections, so the primary experimental evidence for the headline claim is invalid as written. The algorithm description and benchmark comparisons are not obviously wrong, but they do not rescue the path-planning contribution without a corrected formulation and new experiments.","tokens_in":7858,"tokens_out":8879,"duration_ms":95325,"concrete_test":"Implement Eq. (11) literally with m=20 and P=10 on Map 1 from Fig. 2, run any of the four optimizers with population 40 and 200 iterations as specified, and record the argmin's n_O and objective value. If the returned solution has n_O>0, or if the minimal objective value is about 10 rather than about 28m, the penalty does not enforce collision avoidance and Table III cannot support the central claim. For confirmation, recompute with P=10,000; if the reported path lengths change materially, the original penalty was active and the original experiment was optimizing the wrong objective.","verdict_should_be":"REJECT","load_bearing_attack":"The most load-bearing problem is not missing error bars but the stated objective in Eq. (11). For n_O>0, f(X)=P×n_O, and Section III-B sets P=10 on a 20m×20m grid. Feasible start-to-target paths have lengths of roughly 27–31m (Table III; geometric lower bound about 26.9m for cell-center endpoints). Hence any path that intersects a single obstacle cell receives objective value 10, which is strictly smaller than every feasible path's value. The optimizer therefore prefers infeasible shortcuts, and the 'shortest paths' in Table III are not guaranteed to avoid obstacles. This is an internal inconsistency in the formulation, independent of baseline tuning or number of runs; it invalidates the path-planning claim as written. The manuscript even calls P×n_O a 'large penalty' although 10 is smaller than the minimum feasible path length. Unless an unstated boundary constraint silently enforces feasibility, the reported 1.70m–2.00m average advantage is not interpretable as a comparison of obstacle-free paths.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an Improved Grey Wolf Optimizer (IGWO) that augments the standard GWO with an Advanced Cooperative Predation (ACP) mechanism and a Lens Opposition-based Learning (LOBL) strategy, and applies it to UAV shortest path planning. The authors report that IGWO ranks first on benchmark functions F1-F5, F7, and F9-F12, and that paths produced by IGWO are on average 1.70 m, 1.68 m, and 2.00 m shorter than those of GWO, PSO, and WOA, respectively, across four randomly generated 20 m x 20 m maps.","tokens_in":8145,"tokens_out":6669,"duration_ms":69415,"significance":"If the reported results held, the paper would describe a modest but potentially useful incremental contribution: the combination of ACP and LOBL is clearly described and testable, and the application to UAV path planning follows a standard penalty-based formulation. The manuscript's central claims, however, are currently not supported by the evidence as presented. The path-planning objective in Eq. (11) is internally inconsistent with the reported path-length scale, Table III contains no repeated-run statistics, Table II is missing from the manuscript, and several algorithm parameters are unspecified. These issues affect the interpretability of both the benchmark and the path-planning claims, so the paper requires a substantive revision before the proposed method can be fairly evaluated.","major_comments":[{"comment":"The penalty term in Eq. (11) is internally inconsistent with the path lengths reported in Table III. With P=10 and a 20 m x 20 m grid, any path that intersects a single obstacle cell receives an objective value of 10, whereas every feasible path in Table III has length between 27.46 m and 31.39 m, and the geometric lower bound for cell-center endpoints is about 26.9 m. An optimizer minimizing Eq. (11) therefore prefers an infeasible path crossing one obstacle over every feasible path, so the paths in Table III are not guaranteed to be collision-free and the reported 1.70-2.00 m average advantage is not interpretable as a comparison of obstacle-avoiding paths. The sentence in Section II-B that calls P*nO a 'large penalty' is also contradicted by these numbers. The objective must be repaired (e.g., P larger than any possible feasible path length, or an explicit constraint/repair mechanism) and the experiments repeated before the path-planning claim can be assessed.","section":"Section II-B, Eq. (11); Section III-B, Table III"},{"comment":"Table III reports a single path length for each method on each map, with no number of independent runs, no standard deviation, and no statistical test. The central claim that IGWO paths are on average 1.70 m, 1.68 m, and 2.00 m shorter than GWO, PSO, and WOA is therefore based on four point estimates, and the differences (e.g., 0.55 m on Map 1 versus 2.54 m on Map 2) may be within run-to-run variation. The authors should report means and standard deviations over at least 30 runs per map and a paired significance test (or equivalent) before claiming a consistent advantage. In addition, the parameter settings of PSO and WOA are not reported, so it is unclear whether the baselines were fairly configured.","section":"Section III-B, Table III"},{"comment":"Table II, which is supposed to define benchmark functions F1-F13 and their known optimal values, is missing from the manuscript. Since Table I and the benchmark-performance claim depend entirely on the identities of these functions, the reader cannot verify the results or reproduce the experiments. The table must be included and the source of the functions stated.","section":"Section III-A, Table II"},{"comment":"Table I contains precision and summary statistics that need explanation. For example, IGWO on F2 reports avg 2.45E-20 with std 1.23E-39; this is not plausible without a very detailed explanation, because it implies that all 30 independent runs agree to about 40 significant digits. Similarly, zero avg and zero std for F9 and F11 need a note on floating-point precision and convergence criteria. Please state the effective precision, the number of significant digits, and whether the zeros are exact or rounded.","section":"Section III-A, Table I"},{"comment":"The LOBL parameters are underspecified: the lens magnification factor k and the interval endpoints a_j, b_j in Eq. (10) are not given for either the benchmark or the path-planning experiments, and no guidance is provided for choosing them in a 20 m x 20 m grid. Because LOBL is one of the two proposed mechanisms, the missing parameter values make the algorithm non-reproducible and prevent assessing whether IGWO's advantage is due to tuning. Please report the values used for all experiments.","section":"Section II-A, Eq. (10); Section III"}],"minor_comments":[{"comment":"The text refers to 'the objective function, with m=20, P=10, is formulated as shown in Eq. (13)', but the objective function appears in Eq. (11); Eq. (13) does not exist in the manuscript.","section":"Section III-B"},{"comment":"The termination condition is incompletely specified: the text says 'when a convergence threshold is satisfied', but the threshold is never defined, so it is unclear when the algorithm stops.","section":"Section II-B"},{"comment":"The four maps are described as randomly generated, but no random seed or obstacle-cell coordinates are provided, so the path-planning experiments cannot be reproduced from the text alone.","section":"Section III-B, Fig. 2"},{"comment":"Figure 3 is mentioned as a visualization of Table I, but the convergence curves are not discussed or interpreted in the text.","section":"Section III-A, Fig. 3"},{"comment":"The distributions of r3 and r4 are not stated immediately after Eq. (8); the text should explicitly say they are uniformly distributed on (0,1), matching the earlier definitions for r1 and r2.","section":"Section II-A, Eq. (8)"}],"recommendation":"major_revision","confidential_remarks":"This manuscript appears to be an early draft: Table II is missing entirely, Eq. (13) is referenced but absent, and the penalty parameter is inconsistent with the path-length scale. Given the incremental nature of the algorithmic contribution, I recommend requiring a full experimental revision with reproducibility details, proper statistics, and corrected objective formulation before considering publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a GWO variant combining a centroid-aware update (ACP) with lens opposition-based learning (LOBL), tested on CEC-style benchmarks and a 20m x 20m UAV grid. The benchmark results are plausible as far as the missing Table II permits checking, and the idea of using the population centroid to modulate the wolf update is a reasonable tweak. I don't see a fundamental problem with the algorithmic mechanics.\n\nBut the path planning section has a load-bearing flaw that the paper glosses over. The objective in Eq. (11) sets f = P x n_O whenever any obstacle is hit, with P=10. Feasible paths have lengths around 27-31m, so a single obstacle hit yields f=10, which is strictly better than the shortest possible obstacle-free path. That means the optimizer is rewarded for cutting through one obstacle, no matter how long the detour could have been. The authors call P=10 a 'large penalty' but it is smaller than the minimum feasible path length by a factor of about three. Unless there is an unstated constraint that forces waypoints into free cells, Table III is not necessarily comparing collision-free paths, and the 1.70-2.00m average advantage is not interpretable as the paper claims.\n\nThere are other supporting problems: Table II (the benchmark definitions) is missing from the manuscript; Table I has some weird standard deviations (F2 std 1.23E-39 vs avg 2.45E-20 is below machine epsilon relative to the mean); and the four maps are not described with any measure of difficulty or repeated runs, so the path length numbers have no uncertainty. But those are secondary; the objective function issue is the one that breaks the central application.\n\nThe benchmark part might be a legitimate incremental contribution to the GWO-variant literature, and the writing is clear enough that a referee could quickly locate the issues. I would not cite this paper as it stands, and I wouldn't spend a reading group on it unless someone is specifically collecting GWO variants.\n\nFor peer review: I'd send it out, because the flaw is specific, easily demonstrated, and fixable in revision (e.g., f = distance + P x n_O with P >> max feasible length, or a feasibility-handling repair). The referee would need to check whether the corrected objective changes the reported advantages. If the authors fix that, the paper could be an acceptable incremental result.","headline":"The GWO variant is plausible, but the UAV path planning claim collapses on the paper's own objective: with P=10 and feasible paths at ~27-31m, any path hitting one obstacle scores 10 and beats every feasible path.","tokens_in":8694,"tokens_out":4532,"would_cite":false,"duration_ms":45594,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An Improved Grey Wolf Optimizer that adds cooperative predation and lens-based opposition learning plans UAV paths that are, on average, 1.70–2.00 m shorter than GWO, PSO, and WOA across four tested maps.","keywords":["grey wolf optimizer","UAV path planning","shortest path","opposition-based learning","cooperative predation","metaheuristic optimization"],"falsifier":"Rerun each method for 30 independent trials on each of the four maps with identical settings and compare the resulting path-length distributions. The claim would be falsified if IGWO's mean path length is not statistically lower than GWO's, or if the direction of the per-map difference flips across trials.","tokens_in":7656,"feed_emoji":"🐺","tokens_out":11238,"duration_ms":101854,"temperature":0.7,"pith_summary":"This paper tries to establish that an Improved Grey Wolf Optimizer, IGWO, finds shorter obstacle-avoiding paths for a UAV than three baseline metaheuristics. The modification inserts an Advanced Cooperative Predation step that steers search agents with the population centroid and the top three wolves, and a Lens Opposition-based Learning step that replaces agents with a mirrored counterpart after the exploitation phase. On the thirteen benchmark functions tested, IGWO reports the best average value on ten of them, and on four randomly generated 20 m × 20 m maps the paths it produces are on average 1.70 m, 1.68 m, and 2.00 m shorter than those of GWO, PSO, and WOA respectively. If the result holds, it matters because UAV path planning in cluttered environments would gain a simple, training-free improvement of a few meters over existing heuristics.","feed_headline":"Wolf-pack algorithm upgrade shaves ~2 m off UAV flight paths","feed_subtitle":"Adding cooperative predation and lens-based reflection beats GWO, PSO, and WOA on every obstacle map.","key_machinery":"The mechanism is a two-stage modification of the Grey Wolf Optimizer. Advanced Cooperative Predation (ACP) runs before the original exploitation phase: it computes the population centroid $\\vec{X}_{\\text{mean}}(t)=\\frac{1}{n}\\sum_{i=1}^{n}\\vec{X}_i(t)$ and, with a random switch $r_3$, moves each agent toward either the centroid combined with the $\\alpha$ wolf or the centroid combined with the average of the $\\beta$ and delta wolves, using a spiral factor $\\gamma=\\frac{2e^{r(T-t+1)/T}}{4}\\sin(2\\pi r_4)$ to shape the step. Lens Opposition-based Learning (LOBL) runs after exploitation and replaces each agent with its convex-lens mirror image $\\vec{X}(t+1)=\\frac{a_j+b_j}{2}+\\frac{1}{k}\\left(\\frac{a_j+b_j}{2k}-\\vec{X}(t)\\right)$, where $k$ is the lens magnification factor. ACP intensifies local search near the leading wolves; LOBL diversifies the swarm so it can leave unpromising regions.","core_discovery":"The central claim is that IGWO—GWO augmented with ACP before the exploitation phase and LOBL after it—consistently outperforms GWO, PSO, and WOA on the tested benchmark functions and in UAV path planning. On F1–F5, F7, and F9–F12, IGWO reports the best mean objective values, reaching exactly zero mean and standard deviation on F9 and F11. In the path-planning experiments, IGWO yields the shortest path on all four maps, with average reductions of 1.70 m over GWO, 1.68 m over PSO, and 2.00 m over WOA. The authors see these results as evidence that the two added strategies sharpen fine search around the leading wolf while preserving the ability to escape local optima.","pith_inferences":["A decisive follow-up would be repeated runs: running each algorithm 30 times per map and reporting variance would show whether the 1.70–2.00 m average gap is larger than run-to-run noise.","An ablation that disables ACP or LOBL separately would attribute the gains: if removing LOBL eliminates the benchmark wins, the reflection operator is the main driver; if removing ACP eliminates the path-length edge, the centroid-based predation term is what matters.","Since LOBL is a general operator, applying the same two-stage scheme to PSO and WOA could produce similar improvements, which would indicate the gain is not specific to the grey wolf framework.","The ACP update depends on the population centroid, whose behavior in larger maps or higher-dimensional decision spaces is untested; stress-testing on 100 m × 100 m maps or with more waypoints would define where the improvement persists."],"forward_implications":["IGWO ranks first on ten of the thirteen benchmark functions (F1–F5, F7, F9–F12), and on F9 and F11 its reported mean and standard deviation are both exactly zero.","Across all four maps, IGWO's path length is smaller than every baseline's, so the advantage is consistent in sign even though the size varies from 0.55 m to 3.52 m per comparison.","Because IGWO is a training-free metaheuristic with the same swarm size and iteration budget, it can be substituted into existing GWO-based UAV path planners without additional data or tuning.","On F9 and F11 the zero standard deviation means every one of the 30 runs lands on the same optimum, not just a good average."],"supporting_citations":[{"why":"Supplies the base Grey Wolf Optimizer that IGWO augments with ACP and LOBL.","marker":"[17]"},{"why":"Provides the grey wolf cooperative hunting behavior that motivates the Advanced Cooperative Predation update.","marker":"[18]"},{"why":"Introduces the Lens Opposition-based Learning strategy used after the exploitation phase.","marker":"[19]"},{"why":"Defines the benchmark functions used to measure optimization performance.","marker":"[20]"},{"why":"Defines Particle Swarm Optimization, one of the three baseline algorithms IGWO is compared against.","marker":"[21]"},{"why":"Defines Whale Optimization Algorithm, one of the three baseline algorithms IGWO is compared against.","marker":"[22]"}],"fun_headline_variants":["Improved wolf optimizer cuts UAV path length by ~2m","Cooperative predation helps wolf algorithm find shorter UAV routes","Wolf pack upgrade with lens reflection trims UAV paths by ~2m","IGWO: wolf optimizer with advanced tactics shortens UAV flights by ~2m"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim assumes that the four randomly generated 20 m × 20 m maps and the single reported path length per method per map are representative of how GWO, PSO, and WOA would perform in general; any unrepresentative baseline run would make the 1.70–2.00 m average advantage an artifact rather than a genuine property of IGWO.","fun_headline_variants_meta":{"raw":{"variants":["Improved wolf optimizer cuts UAV path length by ~2m","Cooperative predation helps wolf algorithm find shorter UAV routes","Wolf pack upgrade with lens reflection trims UAV paths by ~2m","IGWO: wolf optimizer with advanced tactics shortens UAV flights by ~2m"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000699,"raw_usage":{"total_tokens":3145,"prompt_tokens":919,"completion_tokens":2226,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":2152}},"tokens_in":535,"tokens_out":2226,"duration_ms":17950,"temperature":1.0,"reasoning_tokens":2152,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:57:30.938488+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun each method for 30 independent trials on each of the four maps with identical settings and compare the resulting path-length distributions. The claim would be falsified if IGWO's mean path length is not statistically lower than GWO's, or if the direction of the per-map difference flips across trials.","supporting_citations":[{"cited_title":"Wolf-pack (canis lupus) hunting strategies emerge from simple rules in computational simulations,","cited_arxiv_id":null,"evidence_quote":"Provides the grey wolf cooperative hunting behavior that motivates the Advanced Cooperative Predation update."},{"cited_title":"Opposition-based learning: a new scheme for machine intelligence,","cited_arxiv_id":null,"evidence_quote":"Introduces the Lens Opposition-based Learning strategy used after the exploitation phase."},{"cited_title":"Problem definitions and evaluation criteria for the cec 2005 special session on real-parameter optimization,","cited_arxiv_id":null,"evidence_quote":"Defines the benchmark functions used to measure optimization performance."}],"review_version":1}