Dynamic Dispatching for Time-Sensitive Blood Sample Collection and Delivery

Aliaa Alnaggar; Arash Dehghan; Merve Bodur; Mucahit Cevik

REVIEW 3 major objections 5 minor 42 references

Dynamic Dispatching for Time-Sensitive Blood Sample Collection and Delivery

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A learned dispatch policy raises on-time blood-sample delivery by 1 to 9 percentage points over reactive courier rules.

desk verdict Worth a careful referee, but the training targets in the combined model double-count rewards, so the headline gains over the ablations are likely an artifact. read the letter →

arxiv 2608.04275 v1 pith:CWSNGRIR submitted 2026-08-04 math.OC

Arash Dehghan , Aliaa Alnaggar , Mucahit Cevik , Merve Bodur This is my paper · ORCID

classification math.OC MSC 90C3990C4090B06

keywords bloodsamplecollectionanddeliveryNeurADPvaluefunctionapproximationdynamicvehicleroutingMarkovdecisionprocesshealthcarelogisticsdispatchoptimizationneuralapproximateprogramming

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tackles the daily problem of dispatching couriers to collect blood samples from scattered clinics and return them to a central laboratory before short viability deadlines. The authors claim that a dispatch policy which learns to anticipate future arrivals, rather than react only to the current backlog, can raise the share of sample volume delivered on time by 1 to 9 percentage points relative to myopic baseline policies. The proposed policy, called Combined-NeurADP, learns separate value estimates for vehicles and collection centres and combines them in an integer linear matching program. If the result holds, hospital labs with tight fleets, tight capacities, or tight deadlines could deliver more samples on time and reduce their reliance on costly external couriers.

What carries the argument

The load-bearing mechanism is a dual value-function decomposition: the post-decision value of the whole system is approximated as a sum of vehicle-specific and collection-centre-specific contributions, each produced by its own feedforward neural network trained on entity-level post-decision states. The post-decision state is the system configuration immediately after dispatch decisions are made but before new sample arrivals, so the learned value captures downstream consequences without requiring an expectation over all future arrival paths. At every decision epoch, the dispatcher enumerates feasible single-vehicle routes, scores each route as its immediate collected volume plus the learned downstream value of the vehicle and the visited centres, and solves an integer linear program that assigns each vehicle to one route and each centre to one action. This structure lets the policy trade immediate collection against consolidation while accounting for future arrivals.

What would settle it

A direct test would run the trained Combined-NeurADP policy against the two myopic ILP baselines in a higher-fidelity simulator or a live pilot that uses actual GPS travel times, traffic congestion, and logged sample arrivals from a real laboratory network; if the 1 to 9 percentage-point service-rate advantage disappears in that setting, the central claim would be falsified. A cheaper check on the current simulator would be to replace the review-volume proxy with actual per-centre sample counts and the Dirichlet spatial mixing with the observed empirical spatial distribution; if the performance gaps shrink to zero, the calibration assumptions are carrying the result.

Watch

Extended reading notes

Core claim

The central claim is that Combined-NeurADP, which integrates vehicle-level and centre-level learned value estimates into an ILP dispatch matching model, improves the volume-based service rate over myopic ILP policies by 1 to 9 percentage points in a simulated Greater Toronto Area blood-sample network. The paper further claims that this is the first fully dynamic and stochastic formulation of the blood-sample collection and delivery problem solved with a non-myopic, learning-based approach. Across sensitivity experiments on sample deadline, vehicle capacity, fleet size, and maximum stops per trip, the combined policy attains the highest service rate in almost every setting, with the largest margins under the tightest constraints. In the one setting where routing is trivial (single-centre trips), the centre-only ablation slightly outperforms it, indicating that route-selection flexibility is where the combined architecture earns its advantage. The on-time gains are operationally consequential because samples not served by the primary fleet are assumed to require costly external couriers.

Load-bearing premise

The result depends on the simulated network faithfully representing real blood-sample logistics, in particular travel times computed from straight-line distances at a constant 20 km/h with 5-minute service times and demand generated from a Poisson process calibrated to outpatient data; Appendix A states these are modeling approximations rather than measurements from an operational system, so if real congestion, traffic variability, or demand patterns differ substantially, the measured gains may not transfer.

Editorial extensions

If this is right

At the 90-minute baseline deadline, Combined-NeurADP serves about 93% of sample volume on time, outperforming the ILP-Throughput baseline by 2.23 percentage points and the ILP-Volume baseline by 3.31 percentage points.
The performance advantage grows when resources are scarce: with 3 vehicles the gap over ILP-Volume reaches 6.94 points, and with capacity 10 the gap over Centre-NeurADP reaches 9.41 points.
Relaxing deadlines from 60 to 120 minutes improves service rates for all policies but shrinks the learned policy's advantage, indicating that the main benefit appears in urgent, constrained regimes.
Because samples not served by the primary fleet are assumed to be handled by external couriers at higher cost, the on-time gains translate directly into reduced reliance on that costly backup channel.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same dual supply-demand value decomposition could apply to other persistent-entity logistics settings, such as pharmaceutical distribution or organ transport, where both fleet availability and site-level urgency evolve over the day.
A natural next test is to replace the constant 20 km/h travel-time assumption with time-dependent or stochastic travel times; if the gains persist in that setting, the policy would be substantially closer to deployable practice.
Comparing Combined-NeurADP against a rolling-horizon re-optimization baseline would isolate how much of the reported gain comes from learning a non-myopic value function rather than from simply optimising over a short future window.
The paper evaluates only a single-day horizon with one calibration of demand; retraining and testing on weekly or seasonal arrival patterns would show whether the learned value functions adapt to shifts in the arrival process.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this paper has a good problem, a careful MDP, and a transparent experimental study, but there is a training-target inconsistency in the combined model that double-counts rewards. The headline comparison to the ablations is therefore confounded.

The setup is solid. The MDP formulation with post-decision states is standard and clean, and the MatchingIP is a sensible way to combine entity-level value estimates. The paper also does the right thing in stating that the network and demand data are modeling approximations. The sensitivity analysis around deadline, capacity, fleet size, and route length is useful and consistently reported.

The problem is in Appendix B, Eq. (21). The vehicle target is the route's collected volume plus the target-network value of the vehicle's post-decision state; the centre target is the centre's collected volume plus its target-network value. Summing over all vehicles and centres gives 2R_t + γU_tgt. But the post-decision value in Eq. (19) is the sum of the two networks, and the score in Eq. (20) adds the immediate reward once. So the combined network learns a value that overweights future rewards by an extra factor. The single-vehicle and single-centre ablations are trained consistently, so their comparison to Combined is not apples-to-apples. The reported 1-9 percentage-point gains may simply reflect the double-counted target rather than the benefit of the dual decomposition. That is a load-bearing flaw.

The other weaknesses—no released code or data, missing significance tests—are real but minor next to this. And the paper is honest about the synthetic travel-time and demand models, so I don't count that against it.

I would not cite the current version as evidence for the dual decomposition. But the research question is relevant, the flaw is fixable, and the paper deserves a serious referee who asks the authors to correct the training target and re-run the experiments. If the gains survive that fix, the paper becomes a solid contribution.

Referee Report

3 major / 5 minor

Summary. This paper studies the dynamic dispatch of a courier fleet collecting time-sensitive blood samples from geographically dispersed centres and delivering them to a central laboratory. The authors model the problem as a finite-horizon MDP with stochastic sample arrivals, hard delivery deadlines, vehicle capacity, and multi-stop routes, and propose Combined-NeurADP, which approximates the post-decision value function by a sum of vehicle- and centre-level neural networks and selects joint assignments via an integer linear matching problem (MatchingIP). Experiments on a semi-synthetic Greater Toronto Area testbed compare the policy with two myopic ILP policies and two single-component ablations under variations in deadline, capacity, fleet size, and maximum stops per trip. The headline claim is that Combined-NeurADP improves the on-time served volume by 1 to 9 percentage points over myopic baselines, with larger gains under tighter resource constraints.

Significance. Should the numerical results survive scrutiny, the paper makes a worthwhile methodological and applied contribution: it is, to my knowledge, the first fully dynamic and stochastic formulation of this specimen-collection problem with a non-myopic, learning-based solution, and the dual vehicle-and-centre value decomposition is a natural extension of NeurADP to systems with persistent demand-side state. The MDP is clearly specified, the experimental design covers relevant operating regimes, and the appendix is honest about the synthetic components of the data. The paper does not provide machine-checked proofs or released code/data, so the evaluation rests on the reported experiments; as detailed below, the training-target inconsistency in Appendix B and the absence of statistical significance tests currently prevent me from endorsing the central superiority claim.

major comments (3)

[Appendix B, Eq. (21); Section 5.2.2, Eqs. (19)-(20)] The per-entity TD targets in Eq. (21) assign the full collected volume to both the vehicle network target (q_{a,C'}) and the centre network target (q_t(c)). Because the total value estimate in Eq. (19) is the sum U = Σ_a V̂_Veh + Σ_c V̂_Ctr, summing the per-entity targets over all vehicles and centres yields 2R + γU_tgt rather than R + γU_tgt for the Bellman target of U. Thus U is trained to include an extra copy of every transition reward, and when Eq. (20) adds r(a,C') to γU, the immediate reward is double-counted in the MatchingIP score. The sentence claiming that this attribution is 'consistent with the additive decomposition' conflates allocation (each unit assigned once) with duplication (each unit appears in both the vehicle and the centre sums). This affects the central comparison: the Combined policy's score reweights immediate versus downstream reward relative to the Bellman equation, so the reported gains over myopic baselines and the ablation ordering may be artifacts of the reward duplication rather than evidence for the dual decomposition. The fix is to allocate each transition's reward once across the two components (e.g., train one component on the collected volume and the other on pure downstream value), or to remove r from Eq. (20) and retrain; either way, the experiments must be rerun.
[Section 7, Tables 3-6] The headline comparisons are reported as mean differences without measures of uncertainty. For example, Table 5 (n=5) reports +0.69 pp over Vehicle-NeurADP and +0.65 pp over Centre-NeurADP, and Table 3 (120 min) reports +0.81 pp over Vehicle-NeurADP; with only 20 test days these gaps may not be significant. Since all policies are evaluated on the same 20 sample paths, paired differences can be tested directly. Please report standard errors or confidence intervals for the differences and significance indicators, and also give the benchmark means so the reader can verify the claimed 1-9 pp range.
[Section 6.1 and Appendix A] The simulation uses Haversine distances at a constant 20 km/h with a 5-minute service time, a Dirichlet spatial mixing parameter α=2.0, and review-volume-based prevalence, and Appendix A explicitly labels these as approximations. Because the central claim is a service-rate improvement in a representative operational setting, the absence of sensitivity analysis over these modeling choices leaves open the possibility that the observed ordering is specific to the particular synthetic configuration. Please add robustness experiments (e.g., varying effective speed, service time, α, and the volume distribution) or substantially temper the external-validity language.

minor comments (5)

[Algorithm 1, line 14] The stored experience tuple is written as (S_t, F_t, S^Post_t, S^Post_{t-1}); the fourth element should presumably be the next pre-decision state S_{t+1}. Please correct this, as the ambiguity directly affects the interpretation of the training targets in Eq. (21).
[Appendix B, Table 7] Table 7 lists 'Pricing action exploration Boltzmann sampling, temperature 1' although the model has no pricing decision and Algorithm 1 uses Gaussian score perturbation; this appears to be a leftover from a different application and should be removed or clarified.
[Section 7] Please define SE at first use and clarify that the 'Combined' column reports the mean service rate over the 20 test days, while the '% Over' columns are differences in those means.
[General] The paper does not state code/data availability; providing a repository or a detailed instance generator would materially improve reproducibility.
[Section 5.2.2, Eq. (20)] Equation (20) retains the factor γ although γ=1 throughout the paper; this is harmless, but the notation could be simplified for readability.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the headline service rates are measured from held-out rollouts, not derived from fitted constants; self-citations are contextual rather than load-bearing. The Eq. (21) reward double-count is a potential internal-consistency issue, not a circular derivation.

full rationale

The paper's claimed derivation chain is self-contained: an MDP is formulated, the Bellman equation is rewritten around post-decision states, a dual vehicle-and-centre value approximation is introduced in Eq. (19), actions are scored by Eq. (20), and the resulting MatchingIP is solved at each epoch. The learned value functions are trained on simulated transitions and the reported service rates are computed by rolling out the trained policies on 20 held-out test days; they are not analytic consequences of fitted parameters. The 1-9 percentage-point improvements are empirical outcomes of a simulation, so they do not reduce by construction to the training targets. Self-citations to prior NeurADP work by the same authors appear in the literature review and in statements such as 'in line with prior NeurADP applications', but they are not used as load-bearing evidence for the paper's central derivation, and no uniqueness theorem is imported from the authors' earlier work. The dual decomposition itself is introduced and tested in this paper against external myopic ILP baselines. One substantive concern is not circularity but internal consistency: in Appendix B, summing the per-entity training targets in Eq. (21) gives 2*R_next + gamma*U_tgt, whereas the Bellman-consistent target for the summed value approximation U in Eq. (19) would be R_next + gamma*U_tgt. This could mean the combined network learns an inflated value estimate, potentially affecting the ablation comparison; however, this would be an algorithmic bug or target-alignment error, not a case where the paper's prediction is equivalent to its inputs by definition. Appendix A also openly labels the travel-time, review-volume, and spatial-demand choices as modeling approximations rather than operational measurements, which is a stated limitation rather than a hidden circular step. Overall, no load-bearing circular step was found.

Assumptions & free parameters 11 free parameters · 7 assumptions · 0 invented entities

The ledger shows that the central claim is conditional on a large set of hand-chosen simulation parameters, including fleet size, capacity, deadlines, route limits, travel-speed and demand approximations, and on several domain assumptions about additive value decomposition, exogenous arrivals, and neural-network training convergence. No new physical entities are introduced. The paper is transparent about the modeling approximations, but the experimental conclusions are not backed by released code or data.

free parameters (11)

Baseline fleet size (n) = 4 vehicles (varied: 3, 5)
Chosen for the baseline scenario; fleet scarcity drives the size of the measured improvement.
Baseline vehicle capacity (Q) = 15 volume units (varied: 10, 20)
Capacity is one of the constraints varied in sensitivity analysis; the claim about tight constraints depends on this range.
Baseline sample deadline (tau) = 90 minutes (varied: 60, 120)
Deadline length is a central operational parameter; gains shrink as deadlines relax.
Maximum centres per trip (L) = 2 (varied: 1, 3)
Controls route complexity; at L=1 the centre-only ablation beats the combined policy by 0.66 percentage points.
Decision epoch length (delta) = 5 minutes
Discretization granularity; chosen by the authors, not derived from data.
Terminal clearance period (Delta) = 90 minutes
End-of-day clearing horizon; affects final backlog handling.
Effective courier speed = 20 km/h
Converts straight-line distances to travel times; explicitly an approximation.
Service time per centre = 5 minutes
Added to travel times for handoff and loading; adjusted downward from Grasas et al. (2014).
Dirichlet spatial concentration (alpha) = 2.0
Controls stochastic spatial variation of arrivals across centres; chosen without empirical calibration to Toronto.
Sample volume distribution = Uniform {1,2,3,4,5} units
Assumed volume distribution; not based on measured sample sizes.
Neural network hyperparameters = 3 layers of 300 units, ELU, LR 1e-3, replay exponents 0.6/0.4 to 1.0, target update alpha 1e-3
Algorithmic choices from Table 7; no sensitivity analysis is reported for them, so the headline numbers are conditional on this configuration.

assumptions (7)

standard math Finite-horizon Bellman optimality equation with discount factor gamma=1 characterizes the optimal policy.
Section 4.6, equation (15); standard MDP result, not proven in the paper.
domain assumption The post-decision value function can be represented as a sum of independent per-vehicle and per-centre neural network functions with sufficient accuracy.
Section 5.2.2, equation (19); no approximation error bound is given, and the matching scores are only as good as this decomposition.
domain assumption Future sample arrivals are exogenous and their distribution does not depend on past dispatch decisions.
Section 4.4; standard for this problem, but if arrivals responded to service quality or courier activity, the expectation in the objective would be misspecified.
domain assumption Travel times computed from straight-line distances at constant 20 km/h plus 5-minute service times are representative of real urban courier operations.
Section 6.1 and Appendix A; the authors explicitly call this a modeling approximation, with no congestion or traffic variability.
domain assumption The demand arrival pattern calibrated from outpatient clinic data, with Dirichlet spatial mixing and uniform volumes, matches Toronto collection-centre demand.
Appendix A; calibrated to external data but not to the specific 32 centres, and the review-volume proxy is indirect.
domain assumption Neural network training converges to useful value estimates under the specified hyperparameters and replay schedule.
Appendix B and Algorithm 1; relies on standard deep reinforcement learning practice rather than a convergence proof.
domain assumption Greedy earliest-deadline-first collection at visited centres is an acceptable operational heuristic.
Section 3 and equation (4); the paper does not prove this greedy selection is optimal, and it shapes which samples are collected.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Dynamic Dispatching for Time-Sensitive Blood Sample Collection and Delivery." pith.science (2026). https://pith.science/paper/CWSNGRIR

@misc{pith2026260804275,
  author       = {Pith},
  title        = {Pith review of: Dynamic Dispatching for Time-Sensitive Blood Sample Collection and Delivery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CWSNGRIR}},
  note         = {Machine review of arXiv:2608.04275}
}

read the original abstract

Hospitals and diagnostic laboratories rely on couriers to collect blood samples from geographically dispersed collection centres and deliver them for analysis before their short viability windows expire; late deliveries force costly re-collection and can delay diagnosis. We study the real-time dispatching of such a courier fleet, in which sample requests arrive stochastically at the centres throughout the day and a central dispatcher must repeatedly decide which vehicles to send, which centres each should visit, and whether to collect urgent samples immediately or consolidate them into later trips, subject to hard delivery deadlines and vehicle capacity limits. Unlike static planning models, which fix routes before demand is known, and reactive heuristics, which respond only to the current backlog, our approach anticipates future arrivals when weighing immediate collection against consolidation. We formulate the problem as a Markov decision process and develop a neural approximate dynamic programming framework for centralized dispatch. The method introduces a dual value function decomposition that separately represents vehicle states and collection-centre states through neural networks trained on post-decision states. These learned estimates are integrated through a matching formulation that selects dispatch actions while balancing supply availability, demand urgency, and downstream opportunity cost. Computational experiments on a realistic Greater Toronto Area network compare the proposed policy with myopic baselines and ablation variants. Results show that the proposed dual value function policy raises the share of sample volume delivered on time by 1 to 9 percentage points over myopic baselines, with the largest gains under tight fleet, capacity, deadline, and routing constraints; these on-time gains, in turn, reduce reliance on costly external couriers.

Figures

Figures reproduced from arXiv: 2608.04275 by the authors.

**Figure 2.** Geographic distribution of collection centres (green) and hospital (blue star) in the [PITH_FULL_IMAGE:figures/full_fig_p026_2.png] view at source ↗

**Figure 3.** Expected number of sample arrivals per 5-minute epoch throughout the operating day. [PITH_FULL_IMAGE:figures/full_fig_p026_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Dispatch frequency (a), average travel time (b), and average volume per trip (c) across [PITH_FULL_IMAGE:figures/full_fig_p030_4.png]

**Figure 5.** Figure 5: Coefficient of variation (CV) of centre visits across policies under varying fleet sizes. [PITH_FULL_IMAGE:figures/full_fig_p031_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 39 canonical work pages

[1]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

Neural approximate dynamic programming for on-demand ride-pooling , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

work page
[2]

Transportmetrica B: Transport Dynamics , volume=

An enhanced approximate dynamic programming approach to on-demand ride-Pooling , author=. Transportmetrica B: Transport Dynamics , volume=. 2025 , publisher=

work page 2025
[3]

Conditional Expectation based Value Decomposition for Scalable On-Demand Ride Pooling

Conditional Expectation based Value Decomposition for Scalable On-Demand Ride Pooling , author=. arXiv preprint arXiv:2112.00579 , year=

work page Pith review arXiv
[4]

Proceedings of the 21st International Conference on Autonomous Agents and Multiagent Systems , pages=

Hierarchical Value Decomposition for Effective On-demand Ride-Pooling , author=. Proceedings of the 21st International Conference on Autonomous Agents and Multiagent Systems , pages=

work page
[5]

IEEE Transactions on Intelligent Transportation Systems , volume=

An integrated decomposition and approximate dynamic programming approach for on-demand ride pooling , author=. IEEE Transactions on Intelligent Transportation Systems , volume=. 2019 , publisher=

work page 2019
[6]

Transportation Science , volume=

The restaurant meal delivery problem with ghost kitchens , author=. Transportation Science , volume=. 2025 , publisher=

work page 2025
[7]

Computers & Operations Research , volume=

Same-day delivery with pickup stations and autonomous vehicles , author=. Computers & Operations Research , volume=. 2019 , publisher=

work page 2019
[8]

2007 , publisher=

Approximate Dynamic Programming: Solving the curses of dimensionality , author=. 2007 , publisher=

2007

Show all 42 references

[9]

Dynamic programming and suboptimal control: A survey from

Bertsekas, Dimitri P , journal=. Dynamic programming and suboptimal control: A survey from. 2005 , publisher=

2005
[10]

European Journal of Operational Research , volume=

Solving the dynamic ambulance relocation and dispatching problem using approximate dynamic programming , author=. European Journal of Operational Research , volume=. 2012 , publisher=

2012
[11]

Transportation Science , volume=

A state aggregation approach for stochastic multiperiod last-mile ride-sharing problems , author=. Transportation Science , volume=. 2019 , publisher=

2019
[12]

Computers & Operations Research , volume=

Dynamic lookahead policies for stochastic-dynamic inventory routing in bike sharing systems , author=. Computers & Operations Research , volume=. 2019 , publisher=

2019
[13]

EURO Journal on Transportation and Logistics , volume=

Combining value function approximation and multiple scenario approach for the effective management of ride-hailing services , author=. EURO Journal on Transportation and Logistics , volume=. 2023 , publisher=

2023
[14]

Computers & Operations Research , volume=

Accelerating value function approximations for dynamic dial-a-ride problems via dimensionality reductions , author=. Computers & Operations Research , volume=. 2024 , publisher=

2024
[15]

International Conference on Operations Research and Enterprise Systems , volume=

Dynamic multi-trip vehicle routing with unusual time-windows for the pick-up of blood samples and delivery of medical material , author=. International Conference on Operations Research and Enterprise Systems , volume=. 2016 , organization=

2016
[16]

BMC Health Services Research , volume=

On the improvement of blood sample collection at clinical laboratories , author=. BMC Health Services Research , volume=. 2014 , publisher=

2014
[17]

INFOR: Information Systems and Operational Research , volume=

Routing courier delivery services with urgent demand , author=. INFOR: Information Systems and Operational Research , volume=. 2015 , publisher=

2015
[18]

Journal of the Operational Research Society , volume=

An iterated local search for the biomedical sample transportation problem with multiple and interdependent pickups , author=. Journal of the Operational Research Society , volume=. 2021 , publisher=

2021
[19]

IFAC-PapersOnLine , volume=

A fix-and-optimize variable neighborhood search for the biomedical sample transportation problem , author=. IFAC-PapersOnLine , volume=. 2018 , publisher=

2018
[20]

Computers & Operations Research , volume=

Mathematical programming formulations and metaheuristics for biological sample transportation problems in healthcare , author=. Computers & Operations Research , volume=. 2022 , publisher=

2022
[21]

Computers & Industrial Engineering , volume=

Iterative matheuristic for the biomedical sample transportation problem , author=. Computers & Industrial Engineering , volume=. 2024 , publisher=

2024
[22]

INFORMS Journal on Computing , volume=

An iterative exact algorithm over a time-expanded network for the transportation of biomedical samples , author=. INFORMS Journal on Computing , volume=. 2025 , publisher=

2025
[23]

Computers & Industrial Engineering , volume=

Designing a supply chain network for blood decomposition by utilizing social and environmental factor , author=. Computers & Industrial Engineering , volume=. 2021 , publisher=

2021
[24]

Data , volume=

A dataset of service time and related patient characteristics from an outpatient clinic , author=. Data , volume=. 2023 , publisher=

2023
[25]

Healthcare Courier Services Market - Growth, Trends, and Forecasts , year =
[26]

Blood Needs & Blood Supply , year =
[27]

2025 , howpublished =

2025
[28]

EBioMedicine , volume=

How long can we store blood samples: a systematic review and meta-analysis , author=. EBioMedicine , volume=. 2017 , publisher=

2017
[29]

Biochemia Medica , volume=

Errors within the total laboratory testing process, from test selection to medical decision-making--A review of causes, consequences, surveillance and solutions , author=. Biochemia Medica , volume=. 2020 , publisher=

2020
[30]

, author=

Preanalytical quality improvement: from dream to reality. , author=. Clinical Chemistry & Laboratory Medicine , volume=
[31]

Journal of Patient Safety , volume=

The nature, causes, and clinical impact of errors in the clinical laboratory testing process leading to diagnostic error: a voluntary incident report analysis , author=. Journal of Patient Safety , volume=. 2023 , publisher=

2023
[32]

PSNet , year=

Pre-analytical pitfalls: Missing and mislabeled specimens , author=. PSNet , year=
[33]

MedSpeed , author=

The hidden costs of last-mile logistics failures in Healthcare , url=. MedSpeed , author=. 2025 , month=

2025
[34]

American Journal of Clinical Pathology , volume=

Classifying laboratory incident reports to identify problems that jeopardize patient safety , author=. American Journal of Clinical Pathology , volume=. 2003 , publisher=

2003
[35]

IISE Transactions , pages=

Neural approximate dynamic programming for the ultra-fast order dispatching problem , author=. IISE Transactions , pages=. 2026 , publisher=

2026
[36]

ACM Transactions on Evolutionary Learning , volume=

Dynamic task allocation in intelligent warehouses with hybrid workforce of automated guided vehicles and human pickers , author=. ACM Transactions on Evolutionary Learning , volume=. 2025 , publisher=

2025
[37]

Interfaces , volume=

Approximate dynamic programming captures fleet operations for Schneider National , author=. Interfaces , volume=. 2010 , publisher=

2010
[38]

European Journal of Operational Research , volume=

Approximate dynamic programming for planning a ride-hailing system using autonomous fleets of electric vehicles , author=. European Journal of Operational Research , volume=. 2020 , publisher=

2020
[39]

Transportation Science , volume=

The delivery dispatching problem with time windows for urban consolidation centers , author=. Transportation Science , volume=. 2019 , publisher=

2019
[40]

INFOR: Information Systems and Operational Research , pages=

Joint matching and pricing for crowd-shipping with in-store customers , author=. INFOR: Information Systems and Operational Research , pages=. 2026 , publisher=

2026
[41]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

Optimizing Ride-Pooling Operations with Extended Pickup and Drop-Off Flexibility , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=
[42]

, author=

Value Function Approximation using Multiple Aggregation for Multiattribute Resource Management. , author=. Journal of Machine Learning Research , volume=

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

Neural approximate dynamic programming for on-demand ride-pooling , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

work page

[2] [2]

Transportmetrica B: Transport Dynamics , volume=

An enhanced approximate dynamic programming approach to on-demand ride-Pooling , author=. Transportmetrica B: Transport Dynamics , volume=. 2025 , publisher=

work page 2025

[3] [3]

Conditional Expectation based Value Decomposition for Scalable On-Demand Ride Pooling

Conditional Expectation based Value Decomposition for Scalable On-Demand Ride Pooling , author=. arXiv preprint arXiv:2112.00579 , year=

work page Pith review arXiv

[4] [4]

Proceedings of the 21st International Conference on Autonomous Agents and Multiagent Systems , pages=

Hierarchical Value Decomposition for Effective On-demand Ride-Pooling , author=. Proceedings of the 21st International Conference on Autonomous Agents and Multiagent Systems , pages=

work page

[5] [5]

IEEE Transactions on Intelligent Transportation Systems , volume=

An integrated decomposition and approximate dynamic programming approach for on-demand ride pooling , author=. IEEE Transactions on Intelligent Transportation Systems , volume=. 2019 , publisher=

work page 2019

[6] [6]

Transportation Science , volume=

The restaurant meal delivery problem with ghost kitchens , author=. Transportation Science , volume=. 2025 , publisher=

work page 2025

[7] [7]

Computers & Operations Research , volume=

Same-day delivery with pickup stations and autonomous vehicles , author=. Computers & Operations Research , volume=. 2019 , publisher=

work page 2019

[8] [8]

2007 , publisher=

Approximate Dynamic Programming: Solving the curses of dimensionality , author=. 2007 , publisher=

2007