{"id":"40e13d64-25af-49d4-92f1-59e4bbe43ce1","arxiv_id":"2506.02746","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A reinforcement-learning-controlled adaptive search method finds cheaper pod storage plans than standard heuristics in simulated robotic warehouse tests.","lead":"This paper combines a warehouse-search algorithm with a reinforcement learning agent that decides which search moves to make, and tests it on the problem of where to store robotic shelves after they leave a picking station. The method reaches lower travel costs than most existing rules on a simulated medium-sized warehouse, using a policy trained on a much smaller warehouse.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed medium-instance advantage (DR_ALNS 59.9% vs ALNS 65.24%) rests on one unseeded run of a stochastic search; without repeated runs and variance reporting, the central superiority/transfer claim is not established.","rationale":"The reader's weakest assumption correctly identifies the small-to-medium policy transfer as unsupported, and I agree that the distributional similarity claim is unverified. My independent reading, however, places the more load-bearing problem one step earlier: even the medium-instance result itself, before any transfer interpretation, is a single sample from a randomized algorithm. If DR_ALNS and ALNS were each run repeatedly with different seeds, the 59.9% vs 65.24% gap could shrink, invert, or disappear entirely. That makes the central empirical claim fragile regardless of whether the state representation transfers well. The paper does give a fairly complete algorithmic description and the small-instance result (62.38% vs BIP's 61.85%) is internally consistent, which is genuine supporting evidence that the method is not vacuous. But the distinguishing claim in the strongest claim and in the conclusion is the medium-instance superiority plus transfer, and that claim currently has no statistical support. The reader's CONDITIONAL verdict already captures the need for more evidence, so I would keep the verdict unchanged rather than move to REJECT or UNVERDICTED; the condition is simply that repeated seeded runs and ideally the code/data be supplied.","tokens_in":9768,"tokens_out":6174,"duration_ms":73230,"concrete_test":"Run K=20 independent seeded executions of both ALNS and DR_ALNS on the medium instance, using the same trained PPO checkpoint, the same ALNS hyperparameters, and the same 1,000 ALNS iterations per run. Report mean, min, max, and 95% bootstrap confidence interval for total cost normalized by the random baseline. If the bootstrap CI for the DR_ALNS-minus-ALNS difference includes zero, the claimed 59.9% vs 65.24% advantage and the transfer conclusion are not supported; if the CI excludes zero with comparable widths, the central concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.3 reports DR_ALNS at 59.9% and ALNS at 65.24% on the medium instance, and Section 6 concludes that the policy \"successfully generalized to larger instances without re-training.\" Both claims depend on a single execution of an explicitly stochastic procedure. DestroyRandom selects a uniform random interval (Algorithm 2), RepairRandom assigns uniformly at random (Algorithm 7), simulated annealing accepts with rand() (Section 3.5), and the PPO policy samples from a distribution during inference (Section 4.1). A single trajectory can be arbitrarily lucky, so the reported 5.34-percentage-point gap is not evidence of a true difference unless it exceeds run-to-run variance. No repeated runs, standard deviations, or confidence intervals are reported for any ALNS or DR_ALNS result. The transfer claim is additionally under-supported because it rests on one medium instance only; no comparison is provided between the state-feature distribution encountered during small-instance training and the distribution encountered during medium-instance inference. Section 5.2 tunes ALNS via 50 random parameter combinations on the small instance and then transfers that configuration to the medium instance, so the same single-run weakness applies to both compared methods. Finally, the abstract's claim that DR-ALNS outperforms BIP is contradicted by Figure 1 (BIP 61.85% vs DR_ALNS 62.38%), which further reduces confidence that the reported numbers are robust.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper adapts the DR-ALNS framework (ALNS controlled by a PPO-based deep reinforcement learning agent) to the deterministic Pod Repositioning Problem in robotic mobile fulfillment systems. It defines an MDP in which the agent chooses destroy operators, repair operators, and the degree of destruction, and it introduces several PRP-specific destroy/repair heuristics. The authors evaluate the method on the small and medium benchmark instances of Krenzler et al., reporting cost ratios relative to random placement, and claim that DR-ALNS outperforms BIP, Tetris, and other baselines, and that a policy trained on the small instance generalizes to the medium instance without retraining.","tokens_in":9999,"tokens_out":4435,"duration_ms":44662,"significance":"If the reported advantages are real, the paper would make a useful contribution by showing that learned operator selection can improve ALNS for a deterministic warehouse-storage problem, and the domain-specific heuristics are a reasonable adaptation. The comparison against an exact BIP baseline is a positive feature that gives an independent reference point. However, the central quantitative claims currently rest on single unseeded runs of stochastic algorithms and on hyperparameter selection performed on the same instance used for the final comparison. These issues are load-bearing: the paper's headline result, the medium-instance advantage of DR_ALNS over ALNS, cannot be distinguished from run-to-run noise on the evidence presented.","major_comments":[{"comment":"The reported costs are single values with no repeated runs, standard deviations, confidence intervals, or seed information. The algorithms are explicitly stochastic: DestroyRandom picks a uniform random interval (Algorithm 2), RepairRandom assigns uniformly at random (Algorithm 7), the SA acceptance rule uses rand() (Section 3.5), and the PPO policy samples actions during inference (Section 4.1). The medium-instance gap of 59.9% vs. 65.24% could easily be due to variance. The authors should run multiple independent trials with different random seeds, report dispersion, and ideally provide a statistical test or at least a clear statement of run-to-run spread before claiming superiority or generalization.","section":"Section 5.3, Figures 1 and 2"},{"comment":"The ALNS hyperparameters were selected by testing 50 random combinations on the small instance, and then the same small instance is used for the final comparison in Figure 1. This is a form of selection on the evaluation set, which optimistically biases the ALNS small-instance result and undermines the 'fair comparison' claim. A clean evaluation would tune on a separate validation instance (or use nested cross-validation) and then report performance on a held-out test instance.","section":"Section 5.2 and Figure 1"},{"comment":"The abstract states that DRL-guided ALNS outperforms binary integer programming, but the paper's own Figure 1 text reports BIP at 61.85% and DR_ALNS at 62.38% on the small instance. Thus DR_ALNS does not outperform BIP on that instance. The claim should be qualified as 'comparable or better on some instances' or corrected to reflect the actual data.","section":"Abstract and Section 5.3"},{"comment":"The generalization claim that the policy 'successfully generalized to larger instances without re-training' rests on one medium instance and one inference run. There is no comparison of state-feature distributions between training and inference, no variation of instance parameters, and no repeated medium-instance runs. At minimum, the authors should report multiple medium-instance runs and ideally test on more than one larger instance before drawing a transferability conclusion.","section":"Section 5.3 and Section 6"},{"comment":"The Tetris-inspired repair assigns the pod in Phase 1 to the 'highest-cost location in Pi', which appears to contradict the stated intent of placing the most problematic iterations first in cost-effective locations. If this is a typo (and 'lowest-cost' was intended), the pseudocode must be corrected; if it is literal, the behavior should be explained and justified. As written, the algorithm is ambiguous and affects reproducibility of the reported results.","section":"Algorithm 4, line 7"}],"minor_comments":[{"comment":"The efficiency statement says 'all baseline methods—except for BIP (on the medium instance) and the Genetic Algorithm—complete within one minute for both instance sizes,' but the next sentence reports that ALNS solves the small instance in 103 seconds. This is internally inconsistent; clarify which methods count as baselines and which runtimes are meant.","section":"Section 5.3"},{"comment":"The quantity N in the destroy heuristics is not defined precisely; it is called 'total number of iterations' but the paper also uses ALNS iterations and time steps. Define N as the number of time steps in the planning horizon (or the number of pod-assignment decisions) to avoid ambiguity.","section":"Section 3.2"},{"comment":"The ABC repair heuristic refers to 'second-best' and 'third-best' locations without defining the ranking criterion or tie-breaking rule. This should be specified for reproducibility.","section":"Algorithm 5"},{"comment":"There are several typographical issues, including 'neightborhood' in the keywords, 'illustrating' in the abstract, and inconsistent capitalization of 'Tetris' versus 'tetris-like'. These should be cleaned up.","section":"Abstract and Keywords"},{"comment":"The figures contain only a caption in the text; the numerical values are reported in prose rather than in a table. A table with all reported costs, and ideally the number of runs and standard deviations, would improve readability and verifiability.","section":"Figures 1 and 2"}],"recommendation":"major_revision","confidential_remarks":"The paper has a useful problem adaptation and a reasonable baseline comparison, but the experimental methodology currently cannot support the central claims. The lack of repeated runs and the selection-on-test tuning are fixable within the manuscript's scope, so I recommend major revision rather than rejection. Please ask for a revised version with variance-aware experiments, a separate tuning/validation split, corrected overclaims about BIP, and clarification of the Tetris repair pseudocode."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new part here is the combination of DR-ALNS with PRP-specific destroy and repair operators, plus a transfer claim: a PPO policy trained on the small instance is applied to the medium instance without retraining. That is worth checking out. The paper also gives a complete MDP formulation with pseudocode, and the baselines include BIP, which is an independent lower bound, so the main comparison is not rigged.\n\nThe soft spots are real, though. Every ALNS and DR_ALNS result is a single run with no seeds, no standard deviations, no confidence intervals. The algorithms are explicitly stochastic—random destruction, random repair, simulated annealing acceptance, and PPO sampling all have randomness. One trajectory can be arbitrarily lucky, so the reported 59.9% versus 65.24% gap on the medium instance is not evidence of a true difference unless it clears run-to-run variance. On the small instance DR_ALNS at 62.38% is close to BIP at 61.85%, which is credible for a heuristic, but the abstract's claim that DR-ALNS outperforms BIP is contradicted by Figure 1 and needs a qualifier.\n\nThere is also a tuning issue: ALNS hyperparameters were selected by testing 50 random combinations on the same small instance used for the final comparison. That is a form of selection on the evaluation set, and it inflates ALNS's reported quality relative to what a user would get without insider tuning. The transfer claim rests on one medium instance with no distributional check; it is plausible but not established.\n\nWho is this for? People working on warehouse storage assignment or learning-based control of metaheuristics. The method is sensible, the writing is clear, and the adaptation is honest about building on Reijnen et al. and Krenzler et al. It is not a breakthrough, but it is a reasonable extension with a promising result.\n\nMy recommendation: deserve a serious referee. The paper should go to peer review, and the reviewers should ask for repeated runs with variance reporting, an abstract that does not overclaim against BIP, and ideally some check of state-distribution similarity for the transfer claim. The core idea is publishable if the numbers hold up.","headline":"Solid adaptation of DR-ALNS to pod repositioning, but the headline numbers rest on single runs and the abstract overclaims against BIP.","tokens_in":10602,"tokens_out":1451,"would_cite":false,"duration_ms":15729,"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":"A learned controller for a large-neighborhood search heuristic gives lower pod-storage costs than existing warehouse methods and transfers across instance sizes without retraining.","keywords":["pod repositioning problem","deep reinforcement learning","adaptive large neighborhood search","warehouse logistics","robotic mobile fulfillment systems","storage assignment","operator selection","simulated annealing"],"falsifier":"Run DR_ALNS and plain ALNS on the same medium instance for ten random seeds and compare the cost distributions; if DR_ALNS's costs overlap ALNS's 65.24% rather than centering near 59.9%, the reported gain is a single-trajectory result rather than a stable learned-control effect. Also log the policy's chosen actions on the medium instance and check whether they differ from what ALNS weight-based roulette would choose; if the action distributions match, any improvement would come from the new destroy and repair heuristics, not from the learned controller.","tokens_in":9492,"feed_emoji":"📦","tokens_out":12119,"duration_ms":108811,"temperature":0.7,"pith_summary":"Warehouse robots carry storage pods (movable racks) to picking stations, and after each visit someone must decide where the pod is stored; this paper addresses that Pod Repositioning Problem with a learned controller wrapped around an established metaheuristic. The method, DR-ALNS, is an adaptive large neighborhood search in which a deep reinforcement learning agent chooses at every iteration which destroy operator removes assignments, which repair operator reinserts them, and how much of the solution to destroy. On the small benchmark the result is nearly tied with an exact solver, 62.38% of random-placement cost versus 61.85% for binary integer programming, and on the larger benchmark it has the lowest reported cost, 59.9% versus 65.24% for plain ALNS and 73.8% for the Tetris heuristic. The paper's central claim is that this learned control improves solution quality and that a policy trained only on the small instance works on the larger one without retraining. If true, warehouse storage assignment can be improved by learning to steer a classical search procedure, without giving up the deterministic planning model.","feed_headline":"Learned search beats classic heuristics for warehouse pod storage","feed_subtitle":"A learned controller steers warehouse search to 59.9% of random-placement cost and transfers across scales without retraining.","key_machinery":"The central object is a Markov decision process wrapped around each ALNS iteration. The state is a nine-component vector containing normalized temperature, previous cost change, gap to the best solution found, normalized destroy-operator weights, normalized repair-operator weights, current cost ratio, best cost, cost gap, and progress through the search budget. The action is a single integer decoded into a triple: choose one destroy operator, one repair operator, and a degree of destruction. The domain-specific operators carry the search: Random and High-Cost destruction, and Tetris-inspired, lowest-cost, random, and ABC-priority repair (pods split into A, B, C usage-frequency classes), with feasibility checked by the functions RecallPlaces, IsPlaceFeasible, and FeasibleLocations. The transition applies simulated-annealing acceptance and updates operator weights, and a shaped reward combines cost improvement, new-best bonuses, fluctuation penalties, infeasibility and rejection penalties, and an exploration bonus. This machinery replaces static weight-based roulette with context-dependent decisions, so the search can switch operators and intensity as the annealer cools.","core_discovery":"On the paper's own terms, the discovery is that learning-driven control of adaptive large neighborhood search finds lower-cost pod-to-location plans on both benchmark instances than the deterministic heuristics proposed in the prior formulation of the problem. The reported costs, stated as percentages of a random-placement baseline, are 59.9% on the medium instance (ahead of plain ALNS at 65.24% and the Tetris heuristic at 73.8%) and 62.38% on the small instance (between exact binary integer programming at 61.85% and plain ALNS at 63.68%). The authors further claim that the policy used to select operators was trained exclusively on the 10-location, 1,000-step small instance and applied to the 504-location, 20,000-step medium instance without re-training. The distinct contribution is treating each ALNS iteration as a Markov decision step so that the search's operator choices and destruction intensity adapt to the current temperature, cost gap, and remaining budget, rather than following fixed weight-based roulette selection.","pith_inferences":["The paper leaves implicit that the same policy could carry across warehouse layouts, not just instance sizes, because the state vector does not encode geometry or pod-frequency distributions; this can be tested by training on one layout and evaluating on another.","Only one medium-instance run is reported, so a natural next experiment is to rerun with multiple random seeds and report the spread; the 59.9% figure would be much more load-bearing as a median over seeds.","The reward function actively penalizes fluctuating cost trajectories, so a side effect of the method may be smoother search behavior, which matters if DR_ALNS is embedded in a real-time warehouse control loop where jumpy costs are operationally undesirable."],"forward_implications":["On the medium instance, the reported 59.9% relative cost is below plain ALNS (65.24%) and Tetris (73.8%), so learned operator selection would be the best tested method at larger scale.","A policy trained on 10 pods and 1,000 time steps would transfer to 441 pods and 20,000 time steps, implying that control policies for search operators can be learned in small, cheap experiments and deployed in larger warehouses.","Inference on the medium instance took 400 seconds, versus about 27 minutes for ALNS, so the learned controller would not add runtime burden.","On the small instance DR_ALNS is about half a percentage point behind the exact solver, so it could serve as a near-optimal alternative where exact solving is impractical."],"supporting_citations":[{"why":"Supplies the deterministic PRP model, the two benchmark instances, and the baseline heuristics (Cheapest Place, Fixed Place, BIP, Tetris) against which DR_ALNS is compared.","marker":"[2]"},{"why":"Supplies the DR-ALNS method, deep reinforcement learning controlling operator selection inside ALNS, which the paper adapts to pod repositioning.","marker":"[6]"},{"why":"Supplies the original ALNS adaptive weight-update mechanism that the destroy-repair loop and score-based operator selection build on.","marker":"[8]"}],"fun_headline_variants":["DRL-guided ALNS cuts warehouse pod cost to 60% of baseline","AI learns to pick pod locations, besting classic heuristics","Transfer learning helps DRL pick pod spots without retraining","Deep RL steers warehouse search to lower pod placement cost","Learned adaptive search beats fixed rules for warehouse pods"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The policy was trained only on the small instance, and the transfer claim assumes that after normalization the medium instance's search states look enough like the small instance's that the same operator choices remain good; the paper reports a single medium run and gives no distributional check of that similarity.","fun_headline_variants_meta":{"raw":{"variants":["DRL-guided ALNS cuts warehouse pod cost to 60% of baseline","AI learns to pick pod locations, besting classic heuristics","Transfer learning helps DRL pick pod spots without retraining","Deep RL steers warehouse search to lower pod placement cost","Learned adaptive search beats fixed rules for warehouse pods"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000866,"raw_usage":{"total_tokens":3731,"prompt_tokens":897,"completion_tokens":2834,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":2749}},"tokens_in":513,"tokens_out":2834,"duration_ms":21227,"temperature":1.0,"reasoning_tokens":2749,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:16:50.805264+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DR_ALNS and plain ALNS on the same medium instance for ten random seeds and compare the cost distributions; if DR_ALNS's costs overlap ALNS's 65.24% rather than centering near 59.9%, the reported gain is a single-trajectory result rather than a stable learned-control effect. Also log the policy's chosen actions on the medium instance and check whether they differ from what ALNS weight-based roulette would choose; if the action distributions match, any improvement would come from the new destroy and repair heuristics, not from the learned controller.","supporting_citations":[{"cited_title":"Deterministic Pod Repositioning Problem in Robotic Mobile Fulfillment Systems","cited_arxiv_id":"1810.05514","evidence_quote":"Supplies the deterministic PRP model, the two benchmark instances, and the baseline heuristics (Cheapest Place, Fixed Place, BIP, Tetris) against which DR_ALNS is compared."},{"cited_title":"Artificial Intelligence319 (2024)","cited_arxiv_id":null,"evidence_quote":"Supplies the DR-ALNS method, deep reinforcement learning controlling operator selection inside ALNS, which the paper adapts to pod repositioning."}],"review_version":1}