acceptodds
Under review as a conference paper at ICLR 2027

PoolDFL: Solver-Free Decision-Focused Learning with Augmented Solution Pools

Abstract

Decision-focused learning often requires solving an optimisation problem at every training step. We introduce PoolDFL, which augments a shared pool of training optima with feasible neighbours and keeps the pool fixed throughout training. Neighbour enumeration requires no additional solver calls, and each update evaluates a restricted SPO+ loss through matrix products and maximisation over the pool. This loss recovers the full objective whenever the pool contains a global optimum of its inner optimisation problem. Experiments on shortest path, assignment and knapsack show that augmentation improves decision quality over pools containing only training optima. On shortest path, PoolDFL has a mean regret gap of relative to SPO+ over 35 paired seeds. On assignment, it satisfies a non-inferiority criterion after Holm correction. Gradient diagnostics associate closer alignment with the full objective with smaller regret gaps.

Primary area:
optimization
Keywords:
Decision-focused learning, Predict-then-optimise, Solver-free learning, Solution pools, Neighbourhood augmentation

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_KeZgOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes POOLDFL, a training method that avoids repeatedly invoking optimization solvers in decision-focused learning (DFL) by constructing and expanding a fixed solution pool. The key idea is to approximate the optimization process of the SPO+ loss using optimal training labels and their feasible neighborhoods, with efficient updates implemented through matrix operations.

Critical strengths

None

Critical weaknesses

This paper is very weak. A major issue is that all the links appear to be broken, while the poor formatting further hinders readability. More importantly, the paper presents the enhanced solution pool as its main contribution, but this idea is fairly common and, in my view, is not sufficiently novel to serve as the sole core contribution of a research paper. Finally, the experimental design feels rather ad hoc, with several choices lacking clear justification or a convincing rationale.

Questions for authors

I believe this paper should clearly be rejected, so I have no additional questions.

Reviewer_mncrOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes an approach to decision-focused learning that reduces the cost of training such approaches by reducing the number of solver calls and evaluations necessary. The authors describe their approach and evaluate it empirically.

Critical strengths

The paper considers an interesting problem...

Critical weaknesses

...but is not very accessible for any non-experts in DFL. There is no background section that introduces the concepts; the reader is simply referred to surveys. I did not understand Figure 1 at all; the example needs to be explained in more detail. The content is not very well organised; the description of the method in section 4 contains details specific to the experimental evaluation (unless the proposed method is actually specific to those cases). This makes it harder to understand the details of what is being proposed.

The authors claim that neighbour enumeration does not require any additional solver calls, but presumably the construction of the neighbourhood does? It seems like overall, the cost is still there, it is just incurred at a different time.

Questions for authors

See above.

Reviewer_tunJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PoolDFL constructs a shared pool of training optima, augments it through feasible neighborhood enumeration, and evaluates a restricted SPO+ loss using matrix operations without training-time solver calls. Experiments on shortest path, assignment, and knapsack validate its effectiveness and examine how pool composition, size, and gradient fidelity relate to final regret.

Critical strengths
  1. Simple pool augmentation improves training quality. The method reuses training optima across a shared feasible set and adds candidates through neighborhood enumeration without additional solver calls.
  2. Useful diagnostics distinguish candidate coverage from update quality. Analyses of candidate sources, pool size, scoring rules, and gradient alignment clarify when fixed pools are effective and reveal the limitations of local-neighborhood methods.
Critical weaknesses

The method relies on a fixed feasible set, limiting its applicability. PoolDFL requires candidate solutions to be reusable across instances. When capacities, resources, or graph structures vary across instances, solutions from other instances may no longer be feasible.

Questions for authors
  1. Can the method accommodate instance-dependent capacities, resources, or graph structures? If candidate filtering or repair is required, would the additional overhead undermine its efficiency advantage?
  2. Could you explain why the augmented pool trains more slowly than SPO+ on the assignment task and clarify when to use the base pool, the augmented pool, or direct solver calls?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.