acceptodds
Under review as a conference paper at ICLR 2027

Learning to Consolidate: Candidate-Pool Optimization for Freight Networks

Abstract

We study freight planning in which indivisible shipments can share vehicles on scheduled departures. The large number of feasible paths makes joint optimization difficult, while choosing paths independently can miss opportunities to consolidate freight. We propose a learning-guided method that builds and expands a pool of candidate paths. A neural network scores departures using demand and network information. These scores are converted into shared search costs, under which a path subproblem is solved exactly for each shipment. A restricted mixed-integer master then jointly selects paths and vehicle counts using the original costs and constraints. Loads in the current solution guide the generation of additional paths, which are added to the pool before the master is solved again. On held-out public instances, three neural encoders reduce average cost by 2.54–2.84% and lower median runtime relative to analytic scores in the same search procedure under a common nominal time budget. A separate real-data development case illustrates the method’s computational performance at operational scale.

Primary area:
optimization
Keywords:
Service network design, Freight consolidation, Learning-guided optimization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_gkmhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies freight network optimization, where multiple shipments can share vehicle capacity on scheduled departures. The authors propose WAVECAST, a learning-guided candidate-pool optimization framework.

Critical strengths
  1. The paper presents a learning-guided framework that integrates learned departure preferences with exact path generation.
  2. Experimental results on real data have demonstrated the superiority of the method in large-scale scenarios.
Critical weaknesses
  1. The advantages of the proposed learning-guided strategy are not demonstrated. Although WAVECAST outperforms analytic search, a well-tuned linear scorer achieves comparable solution quality.
  2. The comparisons with scalable optimization methods on large-scale instances are insufficient.
Questions for authors
  1. Could the authors further clarify the benefits of the proposed method, particularly compared with the well-tuned linear scorer?
  2. Could the authors provide more comprehensive comparisons with scalable optimization methods on large-scale instances?
Reviewer_Y38NOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents WAVECAST, a learning-guided candidate-pool optimization framework for freight planning, where each shipment follows a single feasible path while sharing vehicle capacity on scheduled departures. A demand-conditioned encoder predicts departure preferences, which are converted into search charges for an exact shipment-level path oracle. A restricted mixed-integer master then jointly selects candidate paths and vehicle counts under the original costs and constraints. The overall method is a primal heuristic without a global optimality guarantee.

Critical strengths
  1. The experiments are honest and careful. Data splits are properly held out, costs are independently verified, and failures are reported openly

  2. The ablations are thorough. By testing against a simple linear scorer, random scoring, repeated runs of a single model, solver restarts, and perturbed inputs, the authors seem to close off most of the obvious alternative explanations.

Critical weaknesses
  1. My main concern is that this paper is more about pure operational research (OR), although its method has learning components. A large part of this paper is introducing and formulating the freight planning problem, which may not be interesting to most readers of ICLR.

  2. The ablations show that some search components matter more than learning. Across the nine test cases, turning off search components hurts performance by 10–36%, an order of magnitude more than the 0.66–3.40% gained by using a learned field. I doubt that the good empirical performance mostly comes from the OR-style matheuristic machinery, not the ML part.

Questions for authors

Can the proposed learn-to-optimize method be generalized to more problem types?

Reviewer_TmyZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

In this work a freight planning framework with the overall goal to reduce the cost. The proposed approach (WAVECAST) is an optimization approach aided with learned recommendations over shared departures. The approach is evaluated on a public dataset containing 576 instances in 80 source netwrok groups

Critical strengths
  1. The mixed integer programming formulation of the problem is really helpful to an extend.
Critical weaknesses
  1. The introduction section of the manuscript in its present form fails to explain the exact problem being addressed. The introduction section states two specific contributions, however it is not clear what these two contributions are trying to address. It is no clear what the validated cost at the end of section 1 means.
  2. The explanation for the feasibility of a path is not clear.
  3. Please clearly explain what the decision variables are in the MILP formulation.
  4. Section 4.1 doesn't make much sense. There seems to be a set of definitions without explaining what these are and why it matters.
  5. Overall this manuscript is written poorly making it extremely difficult to understand the proposed optimization approach. A significant portion of the work including more experimental evaluations seems to be relegated to the appendices, and are not referenced properly in the main manuscript.
Questions for authors
  1. What is a transportation or ownership mode?
  2. What is a service mode option?
  3. What is Pk in section 3.2?
  4. What does q represent in section 4.1?
  5. What is a charge vector?
  6. What is reference-1602?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.