Learning to Optimize: Joint Routing and Flow Allocation on Sparse Non-Euclidean Networks

Congyuan Ji; Fang He; Haomiao Sun; Xindi Tang

REVIEW 4 major objections 7 minor 1 references

Learning to Optimize: Joint Routing and Flow Allocation on Sparse Non-Euclidean Networks

T0 review · 4 major / 7 minor · reviewed 2026-07-30 · grok-4.5

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

Pith's one-line read A double-channel graph attention policy jointly builds cyclic routes and cargo flows on sparse networks, beating classical and neural solvers at scale in seconds.

desk verdict Solid industrial NCO paper: real sparse/cyclic problem packaging plus a tailored double-channel policy that wins the reported bake-off at scale, but the SOTA-widening claim rests on self-referential batch means with no variance or code. read the letter →

arxiv 2607.23467 v1 pith:CCPBIV35 submitted 2026-07-26 cs.LG

Haomiao Sun , Fang He , Congyuan Ji , Xindi Tang This is my paper · ORCID

classification cs.LG

keywords PickupandDeliveryProblemLearning-BasedOptimizationRoutingFlowAllocationSparseNon-EuclideanNetworksDeepReinforcementLearningGraphAttentionCross-CycleServiceLinerShipping

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

Real liner-style logistics needs one vehicle (or vessel) loop that both chooses which pickups and deliveries to serve and how much cargo to move, on a sparse directed network where straight-line distances do not exist and cargo can finish in the next cycle. Classical optimizers and Euclidean neural routers struggle because feasibility is tight and routing is glued to capacity and due-date logic. This paper introduces Double-Channel Graph Attention (DCGA): it encodes physical reachability and demand-service relations in separate graph channels, then decodes only masked-feasible next stops while a fixed simulator scores the finished loop by profitable flow allocation. On LinerLib-derived instances, once problems pass a modest size, DCGA reports the best batch objectives among time-limited MIP, strong metaheuristics, and adapted neural PDP solvers, with full-batch inference in tens of seconds and a widening lead as size grows. The claim is that structure-aware end-to-end learning is a practical low-latency engine for this joint discrete-continuous problem.

What carries the argument

Double-Channel Graph Attention (DCGA): two GAT channels (network reachability vs. demand-service, linked by logical-physical-logical projection), fused into static node embeddings, then a state-conditioned decoder shaped by a finite delivery-side bias and hard masks (leg, capacity, visit, stop), trained episodically against a simulator that returns the full routing-and-flow objective.

What would settle it

On the same LinerLib-scale batches, if a true joint method (or a policy that also outputs continuous flows) finds substantially better objectives than DCGA’s simulator-scored routes within comparable time, the claimed sufficiency of sequence-only learning plus fixed flow rules would fail.

Watch

Extended reading notes

Core claim

On sparse non-Euclidean pickup-and-delivery instances with capacity coupling, optional fulfillment, and cross-cycle service, an end-to-end RL policy that separates network reachability from demand logic and decodes under hard operational masks plus a deterministic flow simulator achieves state-of-the-art joint objective values beyond a modest scale, with seconds-level inference and a quality advantage that widens as instance size increases.

Load-bearing premise

That letting the neural policy choose only the service sequence, while a fixed rule-based simulator decides fulfilled quantities, loads, and due-date penalties, is good enough to reach high-value joint solutions.

Editorial extensions

If this is right

Medium- and large-scale cyclic routing-and-flow problems on sparse industrial networks can be re-planned in seconds rather than hours.
As network and request size grow, structure-aware learned policies pull further ahead of time-limited MIP and classical local search.
Separating physical topology from OD service logic, plus step-wise hard masks, is necessary to keep neural solvers feasible off the Euclidean plane.
Cross-cycle delivery can be handled without rigid pickup-before-delivery constraints by OD-indexed actions and a soft delivery-side bias.
The same architecture is positioned as a building block for multi-loop scheduling and rolling re-planning under demand shifts.

Reading between the lines

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

If the sequence–simulator split is nearly optimal here, similar hard-mask plus post-hoc continuous allocation patterns may transfer to other capacity-coupled logistics (e.g., empty repositioning, barge networks).
The widening gap with size suggests the main value is constrained exploration efficiency, not just better embeddings—worth testing against pure improvement search seeded by DCGA.
Stability under quantity/price shocks implies the policy may already act as a warm-start generator for exact methods when certificates are required.
Extending the double-channel idea to multi-vessel shared capacity is a natural next stress test of whether channel separation still prevents representational collapse.

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

The useful takeaway is straightforward. They package a realistic liner-shipping-style problem (sparse directed non-Euclidean network, multi-task ports, optional fulfillment, cross-cycle service) as NE-PDP-CCS and ship a double-channel GAT + constraint-masked decoder (DCGA) that, on LinerLib-derived batches from roughly NE-PDP 81 upward, posts the best mean objectives among time-limited MIP, LKH3, ALNS/GA/SA variants, and adapted NCS while finishing full batches in tens of seconds. The gap to the classical methods grows with size in the tables they show.

What is actually new is the problem cut plus the architecture choices that match it: separate network vs demand channels, LPL projection so multi-task physical ports do not fragment the demand graph, OD-indexed actions with a soft delivery-side bias instead of hard precedence, and a composite hard mask (leg/capacity/visit/stop) tied to a deterministic flow simulator. The ablations on masks and single- vs double-channel are clean and point the right way; the demand-perturbation stability check is a plus. Citations cover the right OR and neural-CO baselines, and they give LKH3 and NCS fair adaptations rather than straw men. Math is standard MINLP + CMDP; nothing looks internally contradictory.

Soft spots are real but bounded. The policy only chooses the route sequence; quantities, loads, and due-date penalties come from fixed simulator rules. That is the same scoring environment for every method in Tables 3–4 (LKH3 even gets an LP flow on its sequence), so relative ranking is fair; it just means nobody is solving the true joint continuous optimum. More annoying for the headline claim: tables give only batch means and a gap(%) whose denominator is DCGA itself once it is best, so the “advantage widens” story is partly self-referential. GA2 still beats them on the smallest scale, margins over GA2 later are not huge, and there are no per-instance spreads, win counts, or significance tests. No public code. Inference times omit training amortization. None of that sinks the empirical pattern, but it keeps the SOTA claim conditional on better stats and artifacts.

This is for people who care about neural CO on real sparse logistics graphs or about fast engines inside liner-network design. It deserves a serious referee; I would engage, cite the problem formulation and architecture if I work in the area, and push for code plus instance-level numbers.

Referee Report

4 major / 7 minor

Summary. The paper introduces NE-PDP-CCS, a pickup-and-delivery problem class on sparse directed non-Euclidean networks with capacity-coupled flow allocation, optional fulfillment, soft relative due-date commitments, and cross-cycle cyclic service, motivated by liner shipping. It provides a sequence-space MINLP formulation (Sec. 2), and proposes DCGA: an actor-critic RL solver with a double-channel graph encoder (separating physical reachability from OD demand logic, with a logical-physical-logical projection), a fusion Transformer, a state-conditioned autoregressive decoder, and hard operational masks plus a finite delivery-side bias. Routing is learned; cargo flow, load evolution, and penalties are computed by a deterministic rule-based simulator that evaluates the terminal episodic reward. On LinerLib-derived instances (NE-PDP 61–141), DCGA reportedly achieves the best batch-mean objectives among a time-limited McCormick-relaxed MIP, a decoupled flow model, LKH3, ALNS, GA, SA, and two adapted NCS variants on scales ≥81, with full-batch inference in ~16–20 s, while GA2 beats it at NE-PDP 61. Ablations cover masks, channel count, demand perturbations, and hyperparameters.

Significance. If the results hold, this is a useful contribution to neural combinatorial optimization: a credible demonstration that structure-aware RL policies can outperform time-limited MIP, classical metaheuristics, and adapted NCS solvers on sparse, non-Euclidean, cyclic PDPs at realistic liner-shipping scales, with seconds-level full-batch inference. Strengths worth naming: (i) an explicit MINLP formulation plus an honestly labeled McCormick-relaxed upper-bound benchmark that favors the competitor; (ii) a strong-baseline design in which LKH3 is granted LP-optimal flow on its sequence — a generous evaluation that strengthens the comparison; (iii) mask, channel, perturbation, and hyperparameter ablations; and (iv) a reproducible training protocol described in detail. The NE-PDP-CCS abstraction itself is a plausible benchmark contribution for the NCO and liner-shipping communities. The main caveats limiting significance are statistical (batch means only, self-referential gap metric) and conceptual (the "joint" routing-flow claim is mediated by a fixed rule-based simulator rather than learned or optimized flow).

major comments (4)

[§4.3, Tables 3–4] The central claim — SOTA quality at scale with a widening advantage — is supported only by batch means. No per-instance results, standard deviations, win/loss counts, or significance tests are reported anywhere. Moreover, gap(%) is computed against the best-known objective, which is DCGA itself on every scale ≥81, giving DCGA gap 0.00 by construction and making the 'widening margin' narrative self-referential. The strongest baseline, GA2, in fact beats DCGA at NE-PDP 61 (-12.30 vs -14.03), and the margins at 81/91 are modest relative to batch size. Please report per-instance outcomes (or at least mean ± std, paired win counts, and a significance test against GA2/SA2), and recompute gaps against an external reference (e.g., the relaxed-MIP bound or pooled best-known).
[§3 overview; §3.2.4; App. A] The paper claims to 'jointly optimize cyclic routing and cargo flow allocation,' but the policy only learns the route sequence; fulfilled quantities, load evolution, and tardiness are produced by a fixed rule-based simulator after decoding (§3, App. A). Because all baselines in Tables 3–4 are scored under the same simulator objective — and LKH3 is even granted LP-optimal flow on its sequence (App. G, Table S.5) — the comparative ranking is not asymmetrically inflated. However, the 'joint' claim itself is untested: please apply the same LP-optimal flow evaluation used for LKH3 to DCGA's own decoded routes and report the difference. If rule-based allocation leaves substantial value on the table, the framing and abstract should be qualified accordingly.
[§4.3 narrative; Table S.2] The conclusion that DCGA's advantage 'widens significantly as problem size increases' is confounded by fixed baseline budgets: GA runs 50×100 generations, SA 1500 steps, ALNS fixed depth/width, and LKH3 a per-instance time cap (Table S.2), regardless of instance scale. Metaheuristic quality degrading with size may reflect budget starvation rather than an intrinsic scaling advantage of the learned policy. A concrete check: rerun GA2/SA2/ALNS2 with budgets scaled to instance size (or matched to DCGA's training+inference wall-clock) at two scales and report whether the ordering and margins persist. Training cost of DCGA itself is never reported in Tables 3–4 and should be.
[§4; App. F] No code, trained models, or generated NE-PDP instances are released. The instance-generation procedure (LinerLib Mediterranean subset with controlled OD/demand synthesis) is described only at a high level in §4/App. F, so the main results in Tables 3–4 are not independently verifiable. Given the journal's reproducibility standards, please release the instance generator, test sets, and evaluation code (the simulator in particular, since it defines the objective for all methods).

minor comments (7)

[§4.3, Tables 3–4] The objective is a maximization of profit (Eq. 2), yet all reported objectives are large negative numbers (e.g., -14.03×10^5). Please state explicitly that unmet-demand/tardiness penalties dominate revenue on these instances, and whether all instances are loss-making; otherwise readers will misread the sign convention. The gap formula in §4.3 is also garbled in the text.
[Table 4] The DCGA row appears to be missing its Obj values (only gap and runtime are legible). Please verify the published table.
[§3.2.3–3.2.4, Eqs. (37)–(38)] U_t denotes the raw decoder score in Eq. (37) and is then redefined as the shaped score in Eq. (38). Use distinct symbols (e.g., Ũ and U). Similarly, notation throughout §2 mixes sub/superscript conventions (y_{i,t}, x_{ij,t}, a^p_{r,t}) that are difficult to parse; a notation table would help.
[Table 1] The footnote ('a checkmark indicates at least one cited method considers the feature') makes several rows misleading — e.g., the ALNS row checks TW/DC features that Ropke–Pisinger-style ALNS handles but that differ from the relative due-date commitment here. Consider per-method columns.
[§4.4, Figs. 9–10] The y-axis ('performance gap') of Fig. 9 and the axes of Fig. 10 are never defined; state whether the gap is relative objective difference and against what reference. Batch sizes also vary across scales (600/350/350/350/350/300) without explanation.
[Throughout] Numerous typographical artifacts suggest inadequate proofreading: 'efFectively', 'ibrmulation', 'g；)' for the binomial coefficient in §4.5, '之4*' in Eq. (S.24), 'Optimaliy', and mixed full-width characters in equations throughout §2. A careful pass over the typeset equations is needed.
[§4.1; App. F/G] The adaptations of GA (Jung–Haghani 2000) and SA (Li–Lim 2001) to the NE-PDP-CCS objective, optional fulfillment, and cross-cycle service are not described beyond generic parameters in Table S.2; a short adaptation table analogous to Table S.5 (LKH3) would improve credibility of these baselines.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical RL solver evaluated on external LinerLib-derived instances against third-party baselines under a shared objective.

full rationale

This paper proposes DCGA, trains it with episodic actor-critic on a CMDP, and reports batch mean objectives versus MIP, metaheuristics, and adapted NCS on LinerLib-derived NE-PDP instances. The objective (Eq. 2) is an operational profit definition evaluated by a deterministic flow simulator after route decoding; baselines are scored under the same simulator (LKH3 even gets LP-optimal flow on its sequence). There is no first-principles derivation whose conclusion is forced by its premises, no parameter fitted to data and then re-labeled as a prediction, and no load-bearing uniqueness or ansatz imported via overlapping-author citation. Relative gap(%) to the best-known batch mean is ordinary optimization reporting and does not make the comparative claim tautological—the claim is that DCGA’s measured objectives beat the listed methods, which is falsifiable by those methods. Design choices (double-channel encoder, hard masks, delivery bias, episodic return) shape the policy but do not equate reported performance to the training inputs by construction. Score 0.

Assumptions & free parameters 9 free parameters · 6 assumptions · 4 invented entities

The central empirical claim rests on a CMDP reduction of a nonconvex MINLP, a hand-designed dual-graph inductive bias, a non-learned flow simulator, and many training/decoding hyperparameters chosen for stable RL—not on a parameter-free theorem. Domain fidelity to liner shipping and fairness of adapted baselines are modeling assumptions, not derived facts.

free parameters (9)

Graph/Fusion embedding dimension = 128
Hidden size (default 128) is a capacity knob; sensitivity shows it strongly affects reward.
Graph Encoder attention heads = 8
Head count is tuned; appendix shows clear sensitivity and a preferred setting.
Fusion Encoder layers/heads/FFN dim = layers=4, heads=8, FFN=256
Architectural widths chosen by authors; fusion heads less sensitive but still free.
Delivery-side bias coefficient η_B = 5
Finite bias discouraging unpaired deliveries is hand-set, not derived.
Entropy coefficient ε and advantage coefficient α = ε=0.01, α=1
Policy-gradient regularization weights chosen for exploration/stability.
Actor/critic learning rates and grad clip = 1e-4 / 5e-5, clip=1.0
Adam rates and ℓ2 clip are standard free optimization hyperparameters.
Visit-limit n and termination threshold m in masks = not numerically fixed in main text
Hard mask thresholds that define feasible policy support are design choices.
Route-length bound T and capacity Q / Cmax instance parameters = instance-dependent
Horizon and resource limits shape the CMDP and objective; set per instance family.
Tanh scoring coefficient = 100
Decoder score scale (listed as 100) is a free temperature-like constant.

assumptions (6)

ad hoc to paper Joint NE-PDP-CCS can be optimized by learning only the route sequence while a deterministic rule-based simulator returns feasible cargo flows and the scalar objective.
Stated in §3 overview; underpins all RL training and reported rewards.
domain assumption Sparse reachability, capacity, visit limits, and stop conditions can be enforced exactly by hard decoder masks without needing post-hoc repair.
§3.2.4; ablation Fig. 7 treats masks as necessary for learning.
ad hoc to paper Cross-cycle service is adequately represented by an OD-indexed logical action space plus a finite delivery-side bias rather than hard precedence constraints.
§1–§3 contributions and Eq. (39) bias definition.
domain assumption Mediterranean LinerLib topology with controlled synthetic OD requests is a faithful testbed for industrial sparse non-Euclidean routing-and-flow.
§4 instance construction; Brouer et al. 2014 cited as source.
standard math Episodic Monte Carlo returns with a learned critic baseline and entropy regularization yield a valid policy gradient on the masked feasible set.
§3.4 and Appendix C Boltzmann/entropy surrogate; standard actor-critic assumptions.
domain assumption McCormick relaxation of the bilinear tardiness penalty yields a useful optimistic MIP benchmark under time limits.
Appendix G; used in §4.2 small-batch MIP comparisons.

invented entities (4)

NE-PDP-CCS problem class
purpose: Name and formalize joint cyclic routing, flow allocation, optional fulfillment, and cross-cycle service on sparse non-Euclidean networks.
Defined in §1–§2 MINLP; claimed as broader than prior PDP feature subsets in Table 1.
DCGA (Double-Channel Graph Attention) architecture
purpose: Separate network reachability from demand logic, fuse them, and decode feasible routes with masks.
Core method of §3; performance attributed to this structure via SCGA ablation.
LPL (logical-physical-logical) projection
purpose: Share demand features among multi-task logical nodes that map to the same physical port.
§3.2.1 and Appendix B example; specific inductive bias of this paper.
Composite constraint-informed policy shaping (leg/capacity/visit/stop masks + delivery bias)
purpose: Restrict policy support to operationally valid actions while allowing cross-cycle deliveries.
§3.2.4 Eqs. (38)–(44); ablation shows collapse without masks.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Learning to Optimize: Joint Routing and Flow Allocation on Sparse Non-Euclidean Networks." pith.science (2026). https://pith.science/paper/CCPBIV35

@misc{pith2026260723467,
  author       = {Pith},
  title        = {Pith review of: Learning to Optimize: Joint Routing and Flow Allocation on Sparse Non-Euclidean Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CCPBIV35}},
  note         = {Machine review of arXiv:2607.23467}
}

read the original abstract

We study an integrated pickup-and-delivery problem on sparse, non-Euclidean networks that jointly optimizes cyclic routing, cargo flow allocation, and cross-cycle service. The tight coupling of these operational constraints creates a complex discrete-continuous decision space with highly restricted feasible regions. To overcome these computational challenges, we propose Double-Channel Graph Attention (DCGA), an end-to-end reinforcement learning framework. DCGA isolates network reachability and demand-service logic into separate graph channels and constructs valid routes using a simulator-coupled, constraint-informed decoder. Experiments on LinerLib benchmarks demonstrate that DCGA achieves seconds-level inference and delivers state-of-the-art solution quality on instances beyond a specific scale, with its advantage over existing baselines widening significantly as problem size increases. Supported by extensive stability and ablation analyses, our results demonstrate that this structure-aware learning approach provides an effective, low-latency engine for realistic routing-and-flow optimization.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

1 extracted references · 1 linked inside Pith

[1]

INFOR 32(1):19-31

Bianco L, Mingozzi A, Ricciardelli S, Spadoni M (1994) Exact and heuristic procedures for the traveling salesman problem with precedence constraints, based on dynamic programming. INFOR 32(1):19-31. Brouer BD, Alvarez JF, Plum CE, Pisinger D, Sigurd MM (2014) A base integer programming model and benchmark suite for liner-shipping network design. Transport...

arXiv 1994

Pith tools

Reviewed July 30, 2026 · model on record in the stance chip above.

[1] [1]

INFOR 32(1):19-31

Bianco L, Mingozzi A, Ricciardelli S, Spadoni M (1994) Exact and heuristic procedures for the traveling salesman problem with precedence constraints, based on dynamic programming. INFOR 32(1):19-31. Brouer BD, Alvarez JF, Plum CE, Pisinger D, Sigurd MM (2014) A base integer programming model and benchmark suite for liner-shipping network design. Transport...

arXiv 1994