{"id":"bfb92ad0-6e52-4a1c-be67-57c4a682386c","arxiv_id":"2412.06797","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An Adaptive Large Neighborhood Search heuristic solves the home healthcare routing problem with lunch breaks, matching or improving existing solutions in seconds.","lead":"This paper builds a fast scheduling algorithm that plans home healthcare workers' daily routes while respecting lunch breaks, and wraps it into a decision support system. The algorithm finds optimal or near-optimal solutions for the tested benchmark problems much faster than the existing exact method.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Parameter tuning on 12 instances from the same Liu et al. benchmark used for evaluation, with no held-out validation, leaves the reported optimality rates and 36 improvements potentially overfit; removing the tuning instances from the reported results would test this.","rationale":"The reader identified parameter tuning on the same benchmark as the weakest assumption, and the evidence in the paper supports this concern. Section 4.1 explicitly states that 12 instances from the Liu et al. benchmark were used for tuning, and Section 4.2 reports results on all 168 instances without excluding those 12. Since the algorithm is stochastic and best-of-five results are reported, the evaluation can be optimistically biased. This is the most load-bearing concern because the central claim ('nearly all optimal solutions' and 'better solutions to 36 instances') is a quantitative performance claim on this benchmark, and its transferability to new instances is exactly what the tuning overlap threatens. The 30-patient optimality rate is less sensitive to the tuning set, but the '36 improvements' are mostly on 50- and 100-patient instances, where the tuning instances are concentrated. A simple recomputation without the 12 tuning instances (and using average rather than best-of-five) would settle whether the reported improvements are robust. The verdict remains CONDITIONAL because the concern does not by itself disprove the benchmark results, but it does require additional validation before the general performance claim can be accepted. No code or data are provided, so independent reproduction is not currently possible, further supporting the need for a conditional recommendation.","tokens_in":38016,"tokens_out":14098,"duration_ms":144859,"concrete_test":"Recompute Table 8 and Table 9 after removing the 12 instances used for parameter tuning (Section 4.1) from the evaluation set, and also recompute the improvement count using the average over the five replications (BestAvg) instead of the best (Best). If the number of improved solutions drops below 36, or the average improvement on the remaining 50-patient instances is no longer positive, the reported performance is at least partly an artifact of tuning and best-of-five selection. Report the same statistics on the 12 tuning instances separately for comparison.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 4.1, the parameters (omega, tau_Or, tau_Break) are selected via a full factorial design on 12 randomly chosen 50-patient instances from Liu et al. (2017), and the resulting configuration A0 is then evaluated on all 168 instances in Section 4.2, including those same 12 instances. Because the algorithm is stochastic and the paper reports the best of five replications, any overfitting to the 12 tuning instances can inflate the number of improved solutions and the average gap improvements. The 30-patient optimality claim is less directly affected since tuning used 50-patient instances, but the headline '36 improved solutions' is concentrated in 50- and 100-patient instances, so removing the 12 tuning instances from the 50-patient results could reduce that count. Moreover, no independent validation is provided: the COVID-19 HHCSS experiments in Section 6 have no baseline or optimality comparison. Thus the central claim that the algorithm provides better solutions than Liu et al. on new instances is not yet established; it rests on parameters selected from and evaluated on overlapping data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses the home healthcare scheduling and routing problem with lunch break requirements (HHSRP-LB), following the model of Liu et al. (2017). The authors propose an Adaptive Large Neighborhood Search (ALNS) algorithm with four variants (A0–A3) differing in the use of Or-opt and a novel break-position local search. The method is tested on 168 instances from Liu et al. (2017); the authors report that A0 finds nearly all optimal solutions for 30-patient instances within about 12 seconds on average, obtains better solutions than the published branch-and-price results for 36 instances, visits up to 10 more patients, and is robust with an average coefficient of variation of 0.3. The paper also presents a decision support system (HHCSS) applied to approximately generated COVID-19 patient data from three Turkish cities.","tokens_in":38279,"tokens_out":10059,"duration_ms":80664,"significance":"If the computational claims are correct, the paper would provide a fast, practical heuristic for a realistic HHSRP variant, with transparent reporting (five replications, ANOVA, detailed tables in the appendix). The proposed break-position local search and the DSS are potentially useful practical contributions. However, the central quantitative claims are currently not fully supported: the parameter tuning uses a subset of the same benchmark instances used for evaluation, and there are inconsistencies between the summary tables and the detailed appendix regarding the number of improved solutions. The reported robustness measure is also computed across instances rather than across replications. These issues need to be resolved before the main findings can be accepted.","major_comments":[{"comment":"The parameter tuning in Section 4.1 selects (omega, tau_Or, tau_Break) = (750, 150, 200) using 12 randomly chosen 50-patient instances from Liu et al. (2017), and the evaluation in Section 4.2 uses all 168 instances including those same 12. Because the algorithm is stochastic and reports the best of five replications, any overfitting to the 12 tuning instances can inflate the reported number of improved solutions and the average gaps. This is load-bearing for the abstract's claim of better solutions for 36 instances. The authors should either exclude the 12 tuning instances from the reported results or validate the parameters on a genuinely out-of-sample set.","section":"§4.1 and §4.2"},{"comment":"The variant labeling in the tuning section is internally inconsistent with Section 3.8. The text says the factorial design was run on algorithm A3 'because it consists of all of the three local search heuristics,' but A3 is defined as the configuration with no local search heuristics. It then states that A1 includes only the lunch break local search, while Section 3.8 defines A1 as containing only Or-opt, and A2 as containing only the break heuristic. This swap makes it unclear which of tau_Or and tau_Break was actually tuned for which variant, and it prevents reproduction of the parameter values in Table 5.","section":"§4.1"},{"comment":"The reported counts of improved solutions do not match the detailed appendix. For 50-patient instances, Table 8 reports 11 improvements for A0, but comparing Alg_Best with Liu_Best in Table A2 yields only 8 improvements (C104_50, C204_50, R104_50, R107_50, R108_50, R112_50, R207_50, RC203_50). For 100-patient instances, Table 8 reports 24 improvements, but Table A3 shows at most 21. Table 9 also lists one improvement for RC2_30, while no RC2_30 row in Table A1 has Alg_Best lower than Liu_Best. Because the central contribution is the number of improved best-known solutions, these inconsistencies must be reconciled.","section":"Tables 8/9 vs Tables A1–A3"},{"comment":"The robustness claim is based on a coefficient of variation of 0.3, but Table 9 reports CV_Best as the coefficient of variation of the best objective values across instances within each Solomon class, not across the five replications. That statistic does not measure run-to-run variability of the algorithm; it measures spread across instances with very different scales. With CV_Best values ranging from 0.02 to 0.92, the single average of 0.3 is not a meaningful indicator of algorithmic robustness. The authors should report the coefficient of variation across replications for each instance, or a similar within-instance measure.","section":"§4.2.3 and Abstract"},{"comment":"The sentence 'In the algorithm, the feasibility check is performed only for time windows constraints' appears to contradict the problem constraints, which include caregiver eligibility and qualification requirements (constraints (16) and (17) in Section 2). If the implementation indeed checks only time windows, the solutions may violate skill-matching constraints, which would invalidate the comparison with Liu et al. If eligibility is handled implicitly by the insertion heuristics, this should be stated explicitly. As written, the claim is either incorrect or incomplete and is load-bearing for solution correctness.","section":"§3.6"}],"minor_comments":[{"comment":"The statement 'H0 is rejected because the p-value is 1.0' is statistically backwards: a p-value of 1.0 means the test fails to reject the null hypothesis. The subsequent conclusion that the algorithms are not significantly different is correct, but the wording should be fixed.","section":"§4.2.1"},{"comment":"The noise parameter is denoted as mu in the text of Section 3.3 but as r in Table 5. Please use consistent notation.","section":"§3.3 and Table 5"},{"comment":"There is a duplicated passage in the introduction describing temporal dependency constraints; one of the two copies should be removed.","section":"§1"},{"comment":"The interaction and main effects plots are referenced in Section 4.1 but are not visible in the manuscript; the actual plots should be included in the final version.","section":"Figures 2 and 3"},{"comment":"The statement that the algorithm improved solutions 'on average 6% and 16% in 50- and 100-patient instances' is not directly traceable to Table 9, which reports per-class values rather than an overall average. Please clarify how these averages are computed.","section":"§4.2.3, observation 1"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a practical problem and contains a considerable amount of computational work, but the main claims rest on a benchmark comparison whose integrity is weakened by overlapping tuning/evaluation sets and by inconsistencies between the summary tables and the detailed appendix. The variant-label mix-up in Section 4.1 also affects reproducibility. I recommend inviting a major revision rather than rejecting, because the issues appear fixable: the authors can exclude the tuning instances, correct the variant definitions and counts, and report an appropriate within-replication robustness measure. The DSS section is somewhat disconnected from the benchmark study and would benefit from a clearer evaluation baseline, but this is not the main blocker."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a straightforward ALNS paper for the Liu et al. (2017) home healthcare scheduling problem with lunch breaks. The authors report real, useful results: they solve 30-patient instances to optimality in seconds, and they find better solutions than Liu's branch-and-price on 36 of the 168 benchmark instances. That's a genuine contribution, even if it's incremental.\n\nThe paper does a few things well. The break local search heuristic is a small but sensible addition, and the four algorithm variants let them isolate the effect of each local search component. The computational study is more transparent than most: five replications, ANOVA, and detailed appendix tables. That level of care earns credit.\n\nThe main weakness is the one the stress-test note flags. The three free parameters (omega, tau_Or, tau_Break) are tuned on 12 randomly chosen 50-patient instances from the same Liu et al. set, and then the algorithm is evaluated on all 168 instances, including those 12. Because they report the best of five runs, this overlap can inflate the number of improved solutions and the gap improvements. It's not a fatal flaw, but it does mean the headline '36 improved solutions' needs a robustness check. I'd like to see the 12 tuning instances removed from the results, or at least reported separately. The 30-patient optimality claim is less affected because tuning used 50-patient instances, so that part is probably solid.\n\nTwo smaller issues. No code or data are released, which is a real problem for an ALNS paper whose main contribution is empirical. The COVID-19 DSS section is more of a prototype description; there's no baseline or optimality comparison, so it doesn't add evidence.\n\nOverall, this is a decent applied paper, not a methodological breakthrough. The right audience is practitioners in home healthcare scheduling and researchers using the Liu et al. benchmark. The results are likely real, but the validation overlap and missing code need to be addressed.\n\nI'd send it to a serious referee. The core claim is defensible and useful, and a good referee would ask for the hold-out analysis. Worth engaging with.","headline":"Solid incremental ALNS paper with real benchmark improvements, but the tuning/evaluation overlap and missing code mean the headline results need a validation check.","tokens_in":38816,"tokens_out":3112,"would_cite":true,"duration_ms":31144,"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":"An adaptive large neighborhood search produces near-optimal home-healthcare schedules in seconds.","keywords":["home healthcare","scheduling and routing","lunch break","adaptive large neighborhood search","vehicle routing with time windows","decision support system","metaheuristic","COVID-19 home care"],"falsifier":"Use a holdout set of newly generated instances with different time-window tightness or travel-time distributions, keep the tuned parameters fixed, and compare A0's solutions to branch-and-price or lower bounds; if the optimality rate and the number of improved best-known solutions fall sharply, the reported performance does not generalize.","tokens_in":37858,"feed_emoji":"🏥","tokens_out":6319,"duration_ms":57747,"temperature":0.7,"pith_summary":"The paper tries to show that an Adaptive Large Neighborhood Search algorithm can replace an exact branch-and-price method for daily home healthcare scheduling and routing with lunch breaks. On the 168 benchmark instances used for comparison, the best variant found 105 of the 120 known optima, matched nearly all 30-patient optima in about 12 seconds on average, and improved 36 best-known solutions by up to 36%, serving up to 10 more patients in some classes. If true, this makes near-optimal daily planning practical for home-care agencies and supports a decision support system the authors built for COVID-19 patient visits in Turkish cities.","feed_headline":"ALNS finds near-optimal home-care schedules in 12 seconds","feed_subtitle":"It matches nearly all known optima and improves 36 best solutions, serving up to 10 more patients.","key_machinery":"The load-bearing mechanism is an adaptive large neighborhood search (ALNS), a ruin-and-repair metaheuristic that repeatedly removes patients with one of six removal heuristics and reinserts them with greedy or regret-k insertion heuristics. The lunch break is represented as a special node in every caregiver route, and the insertion procedures explicitly test break-before-service and break-after-service positions; a dedicated break local search then slides the break node to the position that minimizes the route's return time. Variants A0, A1, A2, and A3 selectively include Or-opt and break local search to isolate their contribution, and a simulated-annealing acceptance rule controls diversification.","core_discovery":"The paper's central claim is that its adaptive large neighborhood search solves the home healthcare scheduling and routing problem with lunch breaks almost as well as an exact branch-and-price method but orders of magnitude faster. The A0 variant found 105 of the 120 known optimal solutions, matched nearly all 30-patient optima in about 12 seconds on average, improved 36 best-known solutions by up to 36% in some instance classes, and served up to 10 more patients than the previous best schedules. The authors present this as evidence that a well-constructed metaheuristic with lunch-break-aware insertion is sufficient for practical daily planning.","pith_inferences":["Beyond the paper, the reported gains are concentrated in exactly the instance classes that are hardest for exact methods, such as random locations with narrow time windows, so the method is most valuable where branch-and-price struggles.","Beyond the paper, the statistical equivalence of the four variants suggests the lunch-break positioning built into the insertion heuristics does most of the work, and the dedicated local searches add little; a leaner algorithm might perform identically.","Beyond the paper, because the HHCSS trial uses approximate COVID-19 patient locations derived from public heatmaps, real-world benefit would require validation with actual patient addresses, travel times, and caregiver availability before deployment.","Beyond the paper, the same ruin-and-repair machinery could be extended to multi-period planning, multiple depots, or synchronized visits, as the authors note."],"forward_implications":["A home-care agency can produce daily caregiver itineraries in about 12 seconds for 30 patients and 54 seconds for 100 patients, fast enough for same-day replanning.","The 36 improved best-known solutions, with up to 36% lower cost in the hardest R1_100 class, indicate that exact methods previously left substantial room for improvement on large, tight-window instances.","Serving up to 10 additional patients in some instance classes means the algorithm directly raises service coverage where time windows are narrow.","Because the four algorithm variants are statistically tied on solution quality, the simplest variant without local search may suffice for many practical settings, while adding both local searches is safest on large instances.","The embedded decision support system shows that the algorithm can be wrapped in a map-based tool for scheduling COVID-19-era home care visits."],"supporting_citations":[{"why":"Defines the HHSRP-LB formulation, supplies the 168 benchmark instances, and provides the branch-and-price best-known solutions that the paper compares against.","marker":"Liu et al. (2017)"},{"why":"Provides the ALNS framework, the removal and insertion heuristic repertoire, and default parameter values the paper adopts.","marker":"Ropke and Pisinger (2006a, 2006b)"},{"why":"Supplies the benchmark instance classes (C, R, RC; narrow and wide time windows) on which the test instances are based.","marker":"Solomon (1987)"},{"why":"Contributes the relatedness-based removal heuristic used for removing similar patients.","marker":"Shaw (1997, 1998)"},{"why":"Contributes the Or-opt local search used in algorithm variants A0 and A2.","marker":"Or (1976)"},{"why":"Supplies the simulated-annealing acceptance criterion that controls whether a worse solution is accepted during the search.","marker":"Kirkpatrick (1983)"}],"fun_headline_variants":["Near-optimal home-care routes in just 12 seconds","ALNS improves 36 best home-care schedules by 36%","Home-care scheduling with lunch breaks solved in 12s","Serving 10 more patients: ALNS wins in home-care routing","Decision support cuts home-care scheduling time to seconds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The parameter settings were tuned on 12 of the same 168 benchmark instances used for evaluation, and the paper assumes those settings carry over to all other instances and to new real-world cases.","fun_headline_variants_meta":{"raw":{"variants":["Near-optimal home-care routes in just 12 seconds","ALNS improves 36 best home-care schedules by 36%","Home-care scheduling with lunch breaks solved in 12s","Serving 10 more patients: ALNS wins in home-care routing","Decision support cuts home-care scheduling time to seconds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00051,"raw_usage":{"total_tokens":2493,"prompt_tokens":965,"completion_tokens":1528,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":1444}},"tokens_in":581,"tokens_out":1528,"duration_ms":10952,"temperature":1.0,"reasoning_tokens":1444,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:33:22.279845+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use a holdout set of newly generated instances with different time-window tightness or travel-time distributions, keep the tuned parameters fixed, and compare A0's solutions to branch-and-price or lower bounds; if the optimality rate and the number of improved best-known solutions fall sharply, the reported performance does not generalize.","supporting_citations":[{"cited_title":"D., & Vecchi, M","cited_arxiv_id":null,"evidence_quote":"Supplies the simulated-annealing acceptance criterion that controls whether a worse solution is accepted during the search."}],"review_version":1}