acceptodds
Under review as a conference paper at ICLR 2027

Fusing Optimization Proxies and Reinforcement Learning for Multistage Stochastic Optimization

Abstract

Multistage stochastic programs provide a general framework for sequential decision-making under uncertainty, but their complexity grows exponentially with the number of decision epochs. A popular policy for multistage stochastic optimization solves a two-stage stochastic program at every time step in a rolling-horizon manner, relaxing non-anticipatory constraints for subsequent stages. Solving this sequence of optimization models is computationally expensive, limiting real-time deployment for some classes of applications. This paper presents PROSE (Proxies Reinforcement-learned in Optimization-Structured Environments), which fuses optimization proxies with deep reinforcement learning to obtain a feasible policy for the multistage problem. The optimization model defines the learning problem: at each epoch, the policy takes the same inputs as the two-stage program, including its weighted scenarios. Its action space is the feasible set of the optimization model, enforced by a differentiable repair layer applied to a neural prediction, and its reward is the first-stage cost. The policy is trained by backpropagating the cumulative realized cost through repair and state transitions over complete episodes, end-to-end in a self-supervised manner. On energy dispatch and inventory replenishment case studies, PROSE implements a feasible action at every decision, is 1,600 faster per decision, and typically incurs a lower cumulative realized cost than the rolling two-stage policy.

Primary area:
optimization
Keywords:
multistage stochastic optimization, optimization proxies, reinforcement learning, differentiable layer, sequential decision-making

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_DoeGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes PROSE, a framework for learning policies for multistage stochastic optimization with continuous decisions. At each decision, a NN observes the current state (ztz_t), known optimization parameters, and future scenarios (ηt\eta_t). A state-dependent repair maps its prediction onto the feasible set of actions. Training objective minimizes realized trajectory cost by differentiating through the policy, repair, and known state transitions, without using optimal actions as supervision. The paper evaluates PROSE on stochastic energy dispatch and multi-echelon inventory replenishment. Compared to rolling two-stage stochastic optimization, it reports approximately 4.3 % lower dispatch cost, 1.89% higher inventory profit, and median per-decision speedups of roughly 1600-2100x.

Critical strengths
  1. The combination of trajectory optimization and structured repair is practically motivated. The chain rule accounts for the repair's dependence on the exogenous state. 2, The reported results are encouraging, PROSE reduces total dispatch cost and outperforms the baselines.
Critical weaknesses
  1. Comparison to learning-based methods. The main baseline comparison is with rolling two-stage stochastic optimization. Differentiable predictive control already optimizes neural policies through closed-loop rollouts without any expert labels.
  2. The feasibility guarantee seems to be a slight overclaim. App. B describes an iterative thermal dispatch procedure with a finite transfer cap and an independent feasibility check. The repair might stall or reach its cap with nonzero violations. The current feasibility approach used might not support that more general claim made in the text.
Questions for authors
  1. Can the authors add any learning-based method as a baseline? Just rolling two-stage optimization does not create a sufficient class of baselines. Some examples could be [1], [2]. If the time is insufficient to produce other baseline results, the authors would need to justify why PROSE is better than other learning-based control methods?
  2. What happens if thermal repair finds no improving transfer or reaches the 512-transfer cap with residual overload? Is successful completion guaranteed under stated conditions, or is there a fallback? It would help in understanding how many times the NN policy predicts an action that does not require repair.

[1] Drgoňa et al. (2022), “Learning Stochastic Parametric Differentiable Predictive Control Policies.” [2] Heess et al. (2015), “Learning Continuous Control Policies by Stochastic Value Gradients.”

3 comments
AuthorsOpenReview ↗

Response to Reviewer DoeG

We thank the reviewer for the thoughtful comments and the references to Drgoňa et al. [1] (DPC) and Heess et al. [2] (SVG). W1, W2, ... and Q1, Q2, ... number the Critical Weaknesses and Questions For Authors in order, and G1 to G4 refer to the general response.

[W1, Q1] Missing learning-based baselines such as DPC and SVG. G1 adds a controlled learned baseline, an imitation proxy with PROSE's architecture and repairs, and PROSE beats it in both case studies (Appendix E.2). DPC and SVG are not directly comparable: neither enforces the hard constraints our problems require.

PROSE's constraints are hard, couple the components of each action, and move with the policy's earlier actions, such as an exact power balance within ramp windows set by the previous dispatch. Its closed-form repair enforces them, so every deployed action is certified feasible. DPC penalizes constraint violations and gives only probabilistic guarantees, so its actions can be infeasible (a penalty makes a balance violation small but not zero). DPC has also not been shown to scale to problems of our size: its examples have at most four control inputs and 20-step horizons, against 153 generators and 276 decisions per day in our dispatch case study. SVG has no mechanism for constraint satisfaction and learns a Q-function or dynamics model that PROSE takes as known.

[W2, Q2] Return value if the repair fails. Every implemented action is certified feasible, and the fallback has never been needed (G4). The repair runs at every decision by design, since no raw network output meets an exact power balance. The repair is included in PROSE's reported latency, a median of 17.5 ms per decision against 29.0 s for the two-stage program.

Thank you for the clarifications and updated results. Can you clarify, for a day t, how do you choose the most similar 10 past days? Do you take 32 days following each of the ten similar past days? Do you hard stop at day t?

Thank you for the follow-up.

Scenario selection in inventory. For a decision on day tt, a past day aa is eligible if its demand path, the 32 days aa through a+31a+31, was fully observed before a cutoff. Each eligible day gets a score that adds two terms:

  • How much demand on day a−1a-1 differs from demand on day t−1t-1, the last day observed before the decision. This is the mean squared difference over SKUs, normalized per SKU. The normalization keeps high-volume SKUs from dominating the score.

  • How far apart aa and tt fall in the calendar year, normalized and squared. The calendar wraps at year-end, so late December is close to early January.

Lower scores mean more similar days, and we keep the ten lowest. Appendix C.5 gives the exact scoring rule.

The scenario from a selected day aa is its 32-day demand path, so demand on days aa through a+31a+31 is used as possible demand for days tt through t+31t+31. Each of the ten scenarios has probability 0.1.

Hard stop. Yes. The data are split chronologically into training, validation, and test, and the cutoff depends on the split:

  • In training, the cutoff is day tt, so every path ends by day t−1t-1.

  • In validation and test, the cutoff is the first day of the split, so validation scenarios use only demand from before the validation period, and test scenarios only demand from before the test period.

No scenario ever contains demand from day tt onward, and no validation or test scenario contains demand from its own split. PROSE and the two-stage program use the same scenarios.

Reviewer_1TR8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed PROSE, which replaces a rolling two-stage stochastic optimizer with a neural policy followed by state-dependent repair. The policy is trained on realized trajectory costs by differentiating through repair and system dynamics. Experiments on dispatch and inventory replenishment show better realized outcomes and substantially lower decision latency than the rolling optimization baseline.

Critical strengths
  1. The policy observes the same scenario information as the rolling two-stage program but is trained on realized trajectory costs. This preserves access to forecast uncertainty while allowing learning to account for the downstream consequences of current decisions. The distinction between the look-ahead optimization objective and the closed-loop training objective is well motivated.
  2. The integration of state-dependent repair with rollout differentiation is technically valuable. It allows training to account for both the future cost of an action and its effect on subsequent constraint boundaries. Treating these dependencies together makes the approach useful for sequential problems in which feasibility changes with the policy’s own decisions.
Critical weaknesses
  1. Section 4.4 explains that truncated unrolling improves gradient stability at the cost of dependencies longer than W decision epochs. However, the bias introduced by the stop-gradient operation is not characterized relative to full unrolling. The discussion therefore provides limited guidance on choosing W according to the temporal dependencies of the problem.
  2. The empirical evaluation is insufficient to assess the contribution and reliability of the method. Comparisons are limited to rolling two-stage optimization, without closely related learned baselines or ablations of the principal components. Results across multiple training seeds are also absent. Consequently, the reported gains do not establish which components are necessary or how consistently the improvements can be reproduced.
Questions for authors
  1. HDPO already develops trajectory optimization through known dynamics, and E2ELR establishes self-supervised learning with differentiable repair. What is the principal algorithmic contribution of PROSE beyond integrating these components with scenario-conditioned inputs? Which aspect of the policy formulation or training procedure distinguishes it from these approaches?
  2. Complete recourse is assumed for actions satisfying the original constraints, while training repair may enforce only a subset. What ensures that subsequent transitions and repairs remain well-defined along these relaxed training trajectories?
  3. Under truncated unrolling, are parameters updated at window boundaries or after the complete rollout? Please clarify whether the forward trajectory is generated with fixed parameters, as assumed in Section 4.2.
  4. How are scenario probabilities incorporated into the policy? Is the output invariant to permutations of the weighted scenarios, or does the formulation assume a fixed scenario ordering?
  5. Equation 3 permits time-dependent policies, whereas Equation 7 uses a shared network. Does the policy input include the remaining episode length or another time index that allows it to represent nonstationary finite-horizon decisions?
1 comment
AuthorsOpenReview ↗

Response to Reviewer 1TR8

We thank the reviewer for the comments. W1, W2, ... and Q1, Q2, ... number the Critical Weaknesses and Questions For Authors in order, and G1 to G4 refer to the general response.

[W1] How truncation compares with full unrolling, and how to choose W. Truncation keeps gradients numerically stable, and once W covers how long an action takes to affect later costs, it matches or beats full unrolling in both case studies. Truncation changes only the backward pass. At a window boundary, the state is carried forward and only its derivative is set to zero. In dispatch, full unrolling backpropagates through all 276 decisions of a day, and this long chain makes its gradients less stable: norms reach up to 131, against at most 32 for the truncated windows, so 19.6% of its updates are clipped, against 1.5% and 2.5% at W=12W=12 and 24. The test gaps of W=12W=12, 24, and full unrolling lie within 0.1 points of each other (G3).

Guidance: choose W so that most actions' effects on later costs fall inside their window. In dispatch, the state is the previous dispatch, so 12 to 24 decisions suffice. In inventory, orders arrive up to 27 days later, and a window that ends before an arrival sees only the purchase cost and under-orders: raising W from 16 to 64 lifts the share of orders arriving within the window from 64% to 91%, and test profit relative to the two-stage program from −24.9-24.9\\% to +1.9+1.9\\% (G3). Section 5.4 and Appendix E.4 now state this guidance.

[W2] Missing learned baselines, seeds, and ablations. G1 to G3 add all three. Removing the repair would make every dispatch action and training trajectory infeasible, since a raw neural prediction violates the exact power balance.

[Q1] What PROSE adds beyond HDPO and E2ELR. PROSE's gains rest on two ingredients that neither HDPO nor E2ELR has together: training on the realized multistage cost, and differentiating through a closed-form repair for coupled, state-dependent constraints. HDPO's feasibility layer handles one allocation constraint per node, and E2ELR repairs a deterministic single-period problem, with no coupling between the feasible sets of successive decisions. Our reply to Reviewer CbYD (W4) gives the full comparison.

[Q2] Complete recourse under the relaxed training repair. The training repair relaxes only constraints that do not enter the state transition, so every training transition and repair stays well defined. In dispatch, it relaxes only the line limits, so each repaired dispatch still satisfies every constraint linking consecutive epochs, and overloads only add to the training loss. Deployment enforces every constraint (G4). In inventory, training and deployment use the same allocator, and ordering nothing is always feasible.

[Q3] When parameters update under truncation. Only at window boundaries, never within a window, so each window's forward pass, and with it every gradient, uses fixed parameters, as Section 4.2 assumes.

[Q4] How scenario probabilities and order enter the policy. The dispatch policy uses a fixed scenario order, and the inventory policy is permutation-invariant. In dispatch, the ten nearest-neighbor historical paths, each with probability 0.1, are concatenated into the network input, nearest first, identically in training and deployment, so the network also sees which paths are closest. In inventory, a shared encoder maps each path and averages the encodings.

[Q5] Whether a shared network needs a time input. The input includes no remaining episode length, and none is needed: the shared network makes time-dependent decisions because its inputs change over time. In dispatch, the load, scenarios, and previous dispatch vary with the time of day. In inventory, the inputs include day-of-year features.

Appendix D now states Q2 to Q4.

Reviewer_CbYDOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

PROSE is a framework for a learned policy for multistage stochastic programs. The idea is to replace the rolling two-stage optimization at each epoch with a neural proposal followed by a state-dependent repair into the feasible set. The policy observes the same instance as the two-stage program and is trained end-to-end on the realized closed-loop cost of its own rollouts. Ingredients include pathwise gradients through repair and known dynamics, truncated BPTT, penalties on constraints the training repair does not enforce, and a movement regularizer.

Critical strengths

• Clear problem statement. Defects of the rolling two-stage policy are stated precisely. Assumptions are stated up front and the limitation to continuous decisions is acknowledged. • Feasibility is enforced by construction and certified by an independent audit at every implemented action. • Experimental setup is reasonable.

Critical weaknesses
  1. The baseline's reported cost includes line-overload penalties (Section 5.2; line limits are soft in Appendix A), but PROSE's deployment repair removes all overloads. Section 5.1 still states that every implemented action is feasible. I could not determine how much of the 4.3% gap in Table 1 comes from penalty cost in the baseline.
  2. The only comparator is the two-stage LP. An imitation-trained proxy with the same architecture and repair (as in Ye et al. 2026), or the LP with more scenarios and a longer horizon, is needed to show that the gain comes from closed-loop training and not just from replacing the solver with a network.
  3. In the inventory case, the LP holds prices fixed in its look-ahead while PROSE receives prices as input. The allocator prevents overflow from the current action, but Section 5.4 attributes the zero overflow to learning. How terminal inventory and backlog are valued is not stated. Any one of these could explain the 1.89% margin in Table 3, which comes from a single model on a single trajectory with no seeds or intervals.
  4. The balance repair in Eq. 16 is a closed-form repair of the kind E2ELR introduced, and rollout training follows HDPO and DPC. Unless I am mistaken, TS-DDR (Rosemberg et al. 2025) already trains a parametric policy on the multistage objective. Section 2 identifies the difference as a feasible set that depends on the policy’s earlier actions. The paper does not make the case for why this difference matters. I believe this is important since the paper draws heavily on published ideas.
  5. λmove\lambda_{move} is tuned over three nonzero values (C.2) but a run with λmove=0\lambda_{move} = 0 is not reported. The window WW is swept only in the dispatch case, where the state is the previous dispatch and the window has little room to matter. The inventory case, with lead-time memory, uses W=64W = 64 only.
  6. Transfers in Algorithm 1 decrease the merit and are capped at 512, but this does not guarantee zero overload. The paper does not say what is returned when the cap is reached.
Questions for authors

In addition/related to main issues noted in weaknesses section, following questions come to mind.

  1. How much did the baseline overload lines in test? Report both policies' generation cost without penalties.
  2. Add an imitation-trained proxy with the same architecture and repair
  3. For inventory: how are the ten historical demand paths chosen, what are the lead times relative to the 32-day horizon, and how are terminal inventory and backlog valued?
  4. What solver, hardware, and warm-start settings were used for the baseline?
  5. Can results be reported over multiple seeds?
3 comments
AuthorsOpenReview ↗

Response to Reviewer CbYD

We thank the reviewer for the detailed review. W1, W2, ... and Q1, Q2, ... number the Critical Weaknesses and Questions For Authors in order. The general response answers the imitation baseline (W2, Q2) in G1, seeds (W3, Q5) in G2, ablations (W5) in G3, and the repair's return value (W6) in G4.

[W1, Q1] Penalty cost in the dispatch gap. None of the gap is penalty cost, since the two-stage program never overloads a line. Line slack is zero in all 15,169 test solves, at every scenario node, so Table 1 compares generation-related cost alone.

[W2] Whether a larger two-stage program closes the gap. It does not: it is slower to solve and performs worse. In dispatch, with twice the scenarios and look-ahead, the two-stage program becomes an LP four times as large, takes 322 s per decision instead of 27.5 s, and costs 1.47% more. In inventory, with 15 scenarios (from 10) over 40 decision epochs (from 32), the two-stage program earns 0.71% less (Appendix E.3). More and longer scenarios make the first stage hedge more, positioning more generation and stock ahead of demand. PROSE costs 5.77% less than the larger two-stage program in dispatch and earns 2.62% more in inventory.

[W3] Fairness of the inventory comparison. PROSE has no information or accounting advantage, and its margin holds across five training seeds (G2).

(a) Whether PROSE sees more price information. Both policies see only the current day's prices, with the same state and demand paths. The two-stage program holds them fixed over its look-ahead, since future prices are unobserved.

(b) Why PROSE has zero overflow. It comes from training with the allocator in the loop. The allocator counts shipments in transit against free storage, so no implemented order causes overflow. Sections 5.3 and 5.4 now state this.

(c) How terminal inventory and backlog are valued (Q3). Both policies use the same accounting, and PROSE ends leaner. Purchases are charged when ordered, backlog earns revenue when sold, and the evaluation ends with no salvage value and no charge for remaining backlog. PROSE sells 1,472 more units while buying 7,188 fewer, holds 29% less stock, ends with 36% less backlog and less unsold stock, and still orders 4.3% more over the last 27 days (Appendix E.6).

[W4] Why the differences from E2ELR, HDPO, DPC, and TS-DDR matter. To handle the sequential coupling of a multistage stochastic program at scale, where each decision shapes the costs and feasible sets of later ones, a learned proxy needs two properties together, and none of the four combines them.

(1) Reinforcement learning on the realized multistage cost accounts for how each decision affects later costs, and is why PROSE beats the two-stage program. The rolling two-stage program solves a per-epoch surrogate: a two-stage approximation of the remaining cost over a finite look-ahead. PROSE keeps the same inputs and feasible set but optimizes the realized cost directly. The imitation baseline (G1) isolates this difference: with PROSE's network and repair, learning the two-stage surrogate's actions increases the dispatch cost of the two-stage program by 10.6%, while PROSE decreases it by 4.4%.

(2) Enforcing coupled, state-dependent constraints without a solver makes this training scalable and tractable, since solver-based methods need an optimization solve at every training decision. The repair certifies every deployed action (G4) in a median 17.5 ms per decision, against 228 ms for a quadratic-program projection, which is also hard to differentiate along a trajectory. E2ELR solves a fundamentally different problem (deterministic single-period dispatch) with no coupling between the feasible sets of successive decisions. HDPO's feasibility layer handles one allocation constraint per node, which leaves out coupled constraints such as storage shared across products or dispatch's exact balance and line limits. DPC penalizes constraints, so its actions can be infeasible and its training trajectories can pass through states no feasible policy reaches. TS-DDR solves an optimization problem per scenario at every training iteration, which is costly at scale and requires deriving gradients through every optimization solve.

[Q3] How the demand paths are chosen, and whether the look-ahead covers lead times. The scenarios use only past data, and the 32-day look-ahead covers every lead time. Lead times are at most 10 days per edge and 27 days end-to-end. The ten paths, each with weight 0.1, are the 32 days of demand that followed the ten past days most similar to the given day in the preceding day's demand (normalized per SKU) and the day of year.

[Q4] Solver and hardware settings. In both case studies, HiGHS solves the two-stage program on a single CPU core, and all runs use Intel Xeon Gold 6226 nodes. In dispatch, HiGHS runs through SciPy with default settings on the extensive form, and PROSE uses one NVIDIA Quadro RTX 6000 GPU with 4 CPU cores.

Thanks for your response and follow-up work. I am raising my score due to the rebuttal experiments. It seems to me that every piece exists in prior work. The method needs a hand-built repair per domain. It is a strong workshop-plus paper, but unlikely to change how we think about conceptual advances in multistage stochastic optimization.

Thank you for your engagement and insights, which have helped us improve the paper substantially. We address the two remaining points.

(1) PROSE's contribution is a self-supervised policy, trained end-to-end, that fuses optimization proxies with deep reinforcement learning. To our knowledge, no prior work delivers such a policy with certified feasibility at the scale of our case studies. This policy keeps every decision feasible along long trajectories and makes decisions in milliseconds, fast enough for real-time deployment in real-world operations. Even the rolling two-stage program, an approximation of the multistage problem, has hundreds of thousands of variables and already takes seconds to minutes to solve per decision. In both case studies, PROSE also robustly produces improved decisions over the rolling two-stage program.

(2) For a repair to be fast enough, it must deeply exploit the structure of its problem's constraints, so no single fast repair mechanism can transfer directly to every problem, just as each problem needs its own optimization formulation. PROSE's repair exploits structure further than prior work, enforcing the coupled, state-dependent constraints of sequential decision-making problems without optimization-based projections. More generally, PROSE trains in a self-supervised manner and accepts any piecewise-differentiable map into the feasible set. It serves both energy dispatch and multi-echelon inventory replenishment, two domains with entirely different problem and constraint structures.

Together, these change how multistage stochastic optimization can be approached. Feasible and tractable decision-making for large multistage problems has traditionally required reconstructing and re-solving a stochastic program at every epoch. In contrast, PROSE delivers feasible decisions at real-time speed through a policy learned end-to-end from the optimization model itself, with no need for solver labels, which are expensive to obtain.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers for their comments. W1, W2, ... and Q1, Q2, ... number each reviewer's Critical Weaknesses and Questions For Authors in order. To answer the shared concerns (G1 to G4), we ran four new experiments, now in the paper in blue (Tables 1 and 3, Appendix E): a supervised imitation proxy with PROSE's architecture (G1), more training seeds (G2), ablations of the movement weight and gradient backward window (G3), and a larger two-stage program with more scenarios and a longer look-ahead (W2 of reviewer CbYD). Against the originally reported two-stage program, every PROSE variant trained on the realized cost improves in both case studies, while the imitation proxy and the larger two-stage program both do worse:

& \text{Dispatch cost } \downarrow & \text{Inventory profit } \uparrow \\\\ \hline \text{PROSE} & -4.38\\% & +1.89\\% \\\\ \text{PROSE, mean over training seeds} & -4.10\\% \text{ (3 seeds)} & +1.20\\% \text{ (5 seeds)} \\\\ \text{PROSE, } \lambda_{\mathrm{move}}=0 & -4.38\\% & \text{N/A} \\\\ \text{Supervised imitation proxy (same architecture as PROSE)} & +10.59\\% & -1.98\\% \\\\ \text{Larger two-stage program} & +1.47\\% & -0.71\\% \\\\ \end{array}$$ Entries are changes relative to the originally reported two-stage program (10 scenarios over 12 decision epochs and 10 over 32 in dispatch and inventory, respectively). The larger two-stage program uses 20 over 24 and 15 over 40, and inventory has no movement term. **G1. Missing an imitation-learning baseline** (CbYD W2 and Q2, 1TR8 W2, DoeG W1 and Q1). PROSE's gain comes from training on realized closed-loop cost: the same architecture trained to imitate the two-stage program is worse than that program in both case studies. We trained this proxy (Appendix E.2) by supervised learning on the two-stage program's actions along its trajectory, as in Ye et al. 2026. Unlike PROSE, it also needs the two-stage program's solutions as labels, which take 523.6 solver-hours in dispatch and 5.6 hours in inventory. **G2. Missing results across multiple training seeds** (CbYD W3 and Q5, 1TR8 W2). PROSE's improvement holds across training seeds: every seed beats the two-stage program in both case studies. We retrained PROSE changing only the training seed. In dispatch, the three seeds give $-4.38\\%$, $-4.02\\%$, and $-3.90\\%$, and each is cheaper on 44 to 47 of the 55 test days (15,169 five-minute decisions). In inventory, the five seeds give $+1.89\\%$, $+1.24\\%$, $+0.99\\%$, $+0.97\\%$, and $+0.93\\%$ over the 177-day test trajectory. The first seed in each list is the originally reported model. The revised paper reports the performance across the seeds (Appendix E.1). **G3. Missing ablations of the movement weight and backward window** (CbYD W5, 1TR8 W1 and W2). The movement term is a light regularizer on repair corrections that leaves the dispatch gap unchanged when removed, and the gradient backward window matters when an action takes effect long after it is made, as in inventory. Section 5.2, Table 3, and Appendix E.4 now report both ablations. With $\lambda_{\mathrm{move}}=0$, the dispatch gap is $-4.38\\%$, as with the selected $\lambda_{\mathrm{move}}=10^{-3}$. In dispatch, the validation costs of $W=12$, 24, and full unrolling are within 0.2% of each other, and their test gaps are $-4.30\\%$, $-4.38\\%$, and $-4.28\\%$. In inventory, with lead times up to 27 days, test profit is $-24.9\\%$ at $W=16$, $-6.8\\%$ at $W=32$, $+1.9\\%$ at $W=64$, and $-3.5\\%$ with full unrolling. A window that ends before an order arrives charges its cost but not its sales, and full unrolling's gradients run through the entire training trajectory and can explode or vanish, so a window of proper length, $W=64$ here, does best on both validation and test. **G4. Return value if the repair fails** (CbYD W6, DoeG W2 and Q2). Every action PROSE implemented is feasible, and its repair never failed in any evaluation run. Each action passes an independent audit of every constraint, including line limits. If thermal redispatch stalls or reaches its 512-transfer cap with an overload left, our evaluations would stop with an error, and at deployment PROSE can instead fall back to solving a single-period economic dispatch formulation for that decision. On the test set, the closed-form repair alone certifies 95.5% of decisions, and redispatch handles the other 684 with at most 6 of the 512 allowed transfers. Section 5.1 and Appendix E.5 now state this.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.