acceptodds
Under review as a conference paper at ICLR 2027

SMOP: Self-supervised Multi-period Optimization Proxy for Resource Allocation with Temporal Constraints

Abstract

Multi-period resource allocation problems are typically solved using optimization tools, but solving optimization from scratch can be computationally expensive for large-scale online decision-making. Recent neural-network-based optimization proxies offer substantial speedups, yet they often struggle to preserve feasibility over an entire horizon when global resource-balance constraints interact with inter-temporal constraints: a correction made at the current period may destroy future feasibility. To address this challenge, we exploit the underlying resource–temporal structure and develop a Self-supervised Multi-period Optimization Proxy (SMOP) with three main components: backward–forward feasible-interval propagation, neural-network-based proportional prediction with projection-based feasibility repair, and future-aware forward recovery. We establish full-horizon feasibility guarantees while reducing the nominal computational complexity from cubic to linear in problem size compared with the dense KKT-based differentiable optimization. Experiments on fixed commitment economic dispatch demonstrate that the proposed method achieves millisecond-level inference and produces solutions with zero constraint violations. Across most test scenarios, the resulting feasible solutions also exhibit less than 0.5% optimality gaps without using labels, demonstrating that SMOP can achieve both high computational efficiency and strong solution quality.

Primary area:
optimization
Keywords:
self-supervised, optimization, proxy, multi-period, resource allocation, feasibility, differentiable 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_RdinOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SMOP, a label-free neural proxy for multi-period resource allocation. I do not believe the current work establishes a sufficiently challenging optimization problem or a reliable theoretical contribution to justify acceptance.

  1. The practical difficulty of the underlying power-system problem is overstated. The paper considers fixed-commitment economic dispatch (ED), where unit commitment decisions are given and only continuous generation dispatch variables are optimized. With linear generation costs, generation limits, ramping constraints, and period-wise power balance, this is a structured linear programming (LP) problem. The paper does not address the combinatorial difficulty of unit commitment (UC), where on/off decisions, startup and shutdown decisions, and minimum up/down times introduce integer variables. Nor does it address security-constrained unit commitment (SCUC), which additionally incorporates transmission network constraints and contingency requirements. These are substantially more challenging problems in practical power-system operations. Moreover, the paper evaluates instances with only 10, 50, and 100 generators over 24 periods. Its own results report HiGHS solution times of approximately 7.78, 38.9, and 81.08 ms, respectively. Thus, the proposed speedups of up to 42.58× are measured against an already fast LP solver on relatively small instances. The absolute time savings are only on the order of milliseconds to tens of milliseconds per instance, while the reported neural-network training time ranges from approximately 86 to 400 seconds. The paper does not establish that this trade-off addresses a meaningful bottleneck in actual power-system operations.

  2. Proposition 1 does not establish the claimed full-horizon feasibility guarantee, and Algorithm 4 can violate the original power-balance constraint. See this counterexample (TeX format):

Consider an instance with G=3G=3 generators and T=2T=2 periods. All units are committed, and the parameters are

|p_{g,t}-p_{g,t-1}|\le 0.5,\qquad p_{g,0}=0.5,$$ $$(D_1,D_2)=(1.5,2.7),\qquad R_t=0.$$ The original instance is feasible: $p_{:,1}=(0.5,0.5,0.5)$ and $p_{:,2}=(0.9,0.9,0.9)$ satisfy all constraints. The linear operating costs are irrelevant to this feasibility argument. **Propagated intervals.** At $t=2$, demand balance and the capacity limits of the other two generators imply $p_{g,2}\ge 2.7-2=0.7$. Backward ramp propagation therefore gives $p_{g,1}\ge 0.2$. Repeated interval tightening reaches the fixed point $$[L^B_{g,1},U^B_{g,1}]=[0.2,1],\qquad [L^B_{g,2},U^B_{g,2}]=[0.7,1], \quad g=1,2,3.$$ All intervals are nonempty, and both aggregate demand checks pass. **An allowed prediction and repair.** Choose strictly positive allocation ratios $$\eta_{:,1}=(0.98,0.01,0.01),$$ which are compatible with a softmax output. The corresponding neural prediction is $$\widetilde p_{:,1} =0.2\mathbf{1}+0.8\eta_{:,1} =(0.984,0.208,0.208).$$ The Euclidean box-balance projection onto $\sum_g p_{g,1}=1.5$ has $\lambda=-0.042$ and returns $$\widehat p_{:,1} =\operatorname{clip} (\widetilde p_{:,1}-\lambda,0.2,1) =(1,0.25,0.25).$$ This dispatch satisfies all first-period constraints. However, the second-period reachable upper bounds become $$U^r_{:,2}=(1,0.75,0.75), \qquad \sum_g U^r_{g,2}=2.5<2.7=D_2.$$ Thus, the first-period projection makes the original second-period demand unreachable, even though the original instance is feasible and the propagated intervals pass the aggregate demand checks. Algorithm 4 clips the second-period target to $\widehat D_2=2.5$, producing a power-balance violation of $0.2$ relative to the original demand. Detecting infeasibility instead would also fail to provide a feasible trajectory for the original instance. 3. The experimental results do not resolve the theoretical failure or establish a compelling practical advantage. The experiments support, at most, a fast empirical proxy for a restricted ED formulation, rather than the broader significance implied by the paper's motivation and theoretical claims.
Critical strengths

\

Critical weaknesses

\

Questions for authors

\

3 comments
AuthorsOpenReview ↗

Response to Weakness 1 from Reviewer Rdin

Paper Summary

Weakness 1: This paper proposes SMOP, ...

Response:

MethodMean abs. gap (pp)Max. violationTime (ms/sample)Speedup
SMOP-HC-α, float640.117\mathbf{0.117}<10−9<10^{-9}0.5171.47×1.47\times
SMOP-HC, float640.1220.122<10−9<10^{-9}0.505\mathbf{0.505}1.50×\mathbf{1.50\times}
ConfigurationMean costMean absolute gapMax. ViolationInference time (ms/instance)Speedup
SMOP, batch 120.630.0063$ < 10^{-9}$63.67±5.9463.67\pm5.9416.66×16.66\times
SMOP, batch 820.630.0063$ < 10^{-9}$34.16±1.0934.16\pm1.0931.06×31.06\times
HiGHS Solver20.5000$1060.9 $--

Thank you for these questions. We clarify the scope, computational motivation, and contributions of SMOP below.

(1) SMOP targets a class of multi-period optimization problems with spatial and temporal coupling constraints, not a general solution to UC, SCUC, or arbitrary constrained optimization. Fixed-commitment economic dispatch serves as one application; the present framework does not solve the full mixed-integer unit commitment problem. To demonstrate its broader applicability, we have added two benchmarks—supply-chain optimization and large-scale data-center workload allocation—with formulations and implementations described in Sections 2 and 6 and Appendices F.4 and F.5. For supply-chain optimization, we introduce SMOP-HC and SMOP-HC$-\alpha $ and examine their trade-offs in objective quality, constraint violation, and inference time. For data-center allocation, we evaluate inference time against a conventional LP baseline. Additional ablation studies for both benchmarks are provided in the appendix.

(2) We agree that LP solvers are highly effective for fixed-commitment dispatch, particularly for small instances. Our computational motivation is to reduce the cost of repeatedly solving related problems and to exploit batching particular for large instances. Although conventional solvers can also process independent instances concurrently, neural proxies offer a different opportunity to share computation across a batch. In the DC-Full benchmark in Section 6.4, which contains 1,310 machines over 24 periods, the recorded HiGHS runtime averages 1,060.9 ms per instance, whereas the reported SMOP configuration requires 63.67 ms per instance, corresponding to a 16.67-fold speedup.

(3) SMOP combines neural predictions with structured recovery and establishes feasibility results under explicitly stated conditions. We have provided local correction repair and full-horizon fallback required when those conditions fail, ensuring feasibility. The framework also motivates future integration with neural predictors for discrete decisions. Extending this approach to security-constrained unit commitment would additionally require enforcing commitment logic, network constraints, and feasibility across the full horizon; this remains a direction for future research.

(4) We have added DC3, a self-supervised constrained-optimization baseline that combines equality completion with inequality correction. The results in the table indicate that DC3 has difficulty obtaining feasible solutions efficiently satisfying spatial and temporal constraints in our setting. The comparison strengthens the motivation for recovery mechanisms that explicitly account for temporal coupling.

Comparison of dispatch-learning methods on OR-LIB 10 (mean ± sample standard deviation over three random seeds).

MethodMaximum violation (MW)Absolute gap (%)Inference time (ms/sample)Training time (s)
Supervised57.76±8.5057.76 \pm 8.500.76±0.270.76 \pm 0.270.06±0.002\mathbf{0.06 \pm 0.002}36.59±0.2336.59 \pm 0.23
Supervised + penalties62.24±20.2762.24 \pm 20.270.85±0.220.85 \pm 0.220.07±0.010.07 \pm 0.0136.13±0.80\mathbf{36.13 \pm 0.80}
Supervised + CVXPY(4.36±2.14)×10−6\mathbf{(4.36 \pm 2.14)\times10^{-6}}0.03±0.01\mathbf{0.03 \pm 0.01}2.90±0.152.90 \pm 0.1536.59±0.2336.59 \pm 0.23
Self-supervised + CVXPY(3.54±1.57)×10−6(3.54 \pm 1.57)\times10^{-6}0.62±0.000.62 \pm 0.003.59±0.063.59 \pm 0.06174.4±0.73174.4 \pm 0.73
DC34.35±1.184.35 \pm 1.180.21±0.100.21 \pm 0.103.32±0.073.32 \pm 0.0741.55±1.88\mathbf{41.55 \pm 1.88}
SMOP<10−9\mathbf{<10^{-9}}0.0283±0.0048\mathbf{0.0283 \pm 0.0048}0.683±0.008\mathbf{0.683 \pm 0.008}71.84±0.3671.84 \pm 0.36
AuthorsOpenReview ↗

Response to Weakness 3 from Reviewer Rdin

Weakness 3: Proposition 1 does not establish the claimed full-horizon feasibility guarantee, and Algorithm 4 can violate the original power-balance constraint. See this counterexample (TeX format): Consider an instance with G=3G=3 generators and T=2T=2 periods. All units are committed, and the parameters are 0≤pg,t≤10 \le p_{g,t} \le 1, ∣pg,t−pg,t−1∣≤0.5| p_{g,t}-p_{g,t-1}|\le 0.5, pg,0=0.5p_{g,0}=0.5, (D1,D2)=(1.5,2.7)(D_1,D_2)=(1.5,2.7), Rt=0R_t=0. The original instance is feasible: p:,1=(0.5,0.5,0.5)p_{:,1}=(0.5,0.5,0.5) and p:,2=(0.9,0.9,0.9)p_{:,2}=(0.9,0.9,0.9) satisfy all constraints. The linear operating costs are irrelevant to this feasibility argument. Propagated intervals. At t=2t=2, demand balance and the capacity limits of the other two generators imply pg,2≥2.7−2=0.7p_{g,2}\ge 2.7-2=0.7. Backward ramp propagation therefore gives pg,1≥0.2p_{g,1}\ge 0.2. Repeated interval tightening reaches the fixed point[Lg,1B,Ug,1B]=[0.2,1],[Lg,2B,Ug,2B]=[0.7,1],g=1,2,3.[L^B_{g,1},U^B_{g,1}]=[0.2,1], \qquad [L^B_{g,2},U^B_{g,2}]=[0.7,1], \qquad g=1,2,3. All intervals are nonempty, and both aggregate demand checks pass. An allowed prediction and repair. Choose strictly positive allocation ratios η:,1=(0.98,0.01,0.01)\eta_{:,1}=(0.98,0.01,0.01), which are compatible with a softmax output. The corresponding neural prediction is p~:,1=0.21+0.8η:,1=(0.984,0.208,0.208).\tilde p_{:,1}=0.2\mathbf{1}+0.8\eta_{:,1}=(0.984,0.208,0.208). The Euclidean box-balance projection onto ∑gpg,1=1.5\sum_g p_{g,1}=1.5 has λ=−0.042\lambda=-0.042 and returns p^:,1\hat p_{:,1} = clip (p~:,1−λ,0.2,1)(\tilde p_{:,1}-\lambda,0.2,1) = (1,0.25,0.25)(1,0.25,0.25). This dispatch satisfies all first-period constraints. However, the second-period reachable upper bounds become U:,2r=(1,0.75,0.75),∑gUg,2r=2.5<2.7=D2.U^r_{:,2}=(1,0.75,0.75), \qquad \sum_g U^r_{g,2}=2.5<2.7=D_2. Thus, the first-period projection makes the original second-period demand unreachable, even though the original instance is feasible and the propagated intervals pass the aggregate demand checks. Algorithm 4 clips the second-period target to D^2=2.5\hat D_2=2.5, producing a power-balance violation of 0.20.2 relative to the original demand. Detecting infeasibility instead would also fail to provide a feasible trajectory for the original instance.

Response: Thank you for providing this counterexample. We have revised Section 4 to introduce an explicit next-period feasibility check and a local feasibility repair. The updated proposition and accompanying remark clarify the scope of these safeguards. (1) Next-period feasibility check and local correction. We have added Corollary 1 in Section 4, which introduces an explicit next-period feasibility check and a local repair. We first adjust the neural dispatch proposal to obtain pt(0)p_t^{(0)}. SMOP then computes the next-period bounds induced by $ p_t^{(0)} $, i.e., Lg,t+1r(pt(0))L^r_{g,t+1}(p_t^{(0)}) and Lg,t+1r(pt(0))L^r_{g,t+1}(p_t^{(0)}). It accepts this tentative dispatch only if every next-period interval is nonempty and ∑gLg,t+1r(pt(0))≤Dt+1≤∑gUg,t+1r(pt(0))\sum_g L^r_{g,t+1}(p_t^{(0)})\le D_{t+1}\le\sum_g U^r_{g,t+1}(p_t^{(0)}). If either condition fails, SMOP seeks a replacement dispatch by solving min⁡x\min_x $ ||x-\tilde{p_t}||^2$ s.t. ∑gxg=Dt\sum_g x_g=D_t, Lg,tr≤xg≤Ug,trL^r_{g,t}\leq x_g\leq U^r_{g,t}, Lg,t+1r(x)≤Ug,t+1r(x)L^r_{g,t+1}(x)\le U^r_{g,t+1}(x), ∑gLg,t+1r(pt(0))≤Dt+1≤∑gUg,t+1r(pt(0))\sum_g L^r_{g,t+1}(p_t^{(0)})\le D_{t+1}\le\sum_g U^r_{g,t+1}(p_t^{(0)}). Under the assumptions of Proposition 1, the final two constraints ensure that a next-period dispatch satisfy all the constraints. The first-period dispatch in the counterexample violates this acceptance criterion and therefore cannot be accepted unchanged. Once the bounds are available, the check requires O(G)O(G) operations; the local optimization is invoked only when the check fails.

(2) Full-horizon fallback. One-step extendability does not, by itself, establish full-horizon feasibility. If the local correction problem is infeasible, SMOP falls back to the original full-horizon LP, allowing earlier tentative dispatches to be revised. Provided that this LP is feasible and the solver returns a solution that passes the stated constraint checks, the fallback supplies a feasible full-horizon trajectory.

(3) Training regularization. The bound-feasibility loss, Lbounds\mathcal L_{\mathrm{bounds}}, complements these safeguards by penalizing demand outside the dynamically reachable interval. Because future bounds depend on earlier dispatches, this loss provides gradients that encourage decisions with feasible continuations, as explained in Proposition 4. This regularization can reduce the need for correction in training.

Our updated experiments provide empirical evidence under the evaluated settings. For example, all 301 DC-Full validation instances passed the independent constraint audit at a tolerance of 10−610^{-6}, with a maximum violation across all audited constraints of 5.60×10−75.60\times10^{-7}cores. These results support the effectiveness of the evaluated recovery pipeline without establishing an unconditional feasibility guarantee.

AuthorsOpenReview ↗

Response to Weakness 2 and Weakness 4 from Reviewer Rdin

Weakness 2: Moreover, the paper evaluates instances with only 10, 50, and 100 generators over 24 periods. Its own results report HiGHS solution times of approximately 7.78, 38.9, and 81.08 ms, respectively. Thus, the proposed speedups of up to 42.58× are measured against an already fast LP solver on relatively small instances. The absolute time savings are only on the order of milliseconds to tens of milliseconds per instance, while the reported neural-network training time ranges from approximately 86 to 400 seconds. The paper does not establish that this trade-off addresses a meaningful bottleneck in actual power-system operations.

Response: Thanks for the suggestions. (1) We have added another benchmark of 1310 unit with 17,032 input dimensions mainly in Section 6.4, which taks about 1060 ms for LP solver to find solutions. Experiment demonstrates that SMOP can process all 1,310 machines and that batched inference is substantially faster than the stored LP solve-time baseline under the reported timing scopes. The fastest measured sweep configuration achieves 57.18$ \times $speedup, as shown in Table 4 of Section 6.4. (2) Training time might be long but occurs offline, and the trained model is able to solve a number of instances at a time. Only the inference time matters in reality when comparing with solvers. (3) The benchmark with 1310 units demonstrate the capability of SMOP for practical applications in power grids when hundreds and thousands of units are considered.

Weakness 4: The experimental results do not resolve the theoretical failure or establish a compelling practical advantage. The experiments support, at most, a fast empirical proxy for a restricted ED formulation, rather than the broader significance implied by the paper’s motivation and theoretical claims.

Response: Thank you for the feedback.

  1. Feasibility safeguards. We introduce three mitigation strategies to address cases in which the conditions of Proposition 1 are not satisfied. These safeguards support feasibility recovery, with the updated experiments demonstrating small constraint violations within the stated numerical tolerances. We distinguish this empirical evidence from an unconditional theoretical guarantee.

  2. Applicability beyond economic dispatch. We extend SMOP to supply-chain optimization and large-scale data-center workload allocation. These benchmarks demonstrate its adaptability across domains and its computational advantages for large-scale, repeated optimization, particularly with batched inference.

  3. Broader significance. SMOP provides a foundation for efficient neural proxies for multi-period optimization, combining learned predictions with structured feasibility recovery while maintaining small optimality gaps. Its design also suggests promising directions for mixed-integer multi-period optimization and feasibility-aware combinatorial optimization.

Reviewer_jDHxOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SMOP, a self-supervised optimization proxy for multi-period resource allocation with temporal constraints. The framework combines backward–forward feasible-interval propagation, neural network-based allocation prediction, and sequential projection-based repair to maintain feasibility across the planning horizon. Experiments on economic dispatch benchmarks demonstrate low constraint violations, small optimality gaps, and efficient inference.

However, the paper does not sufficiently discuss relevant literature on learning-based feasibility handling. The experimental evaluation is also limited in generalizability, with insufficient analysis and potentially unfair baseline comparisons. Therefore, I lean toward rejection.

Critical strengths

Structure-aware feasibility handling. Exploiting problem-specific structure to simplify feasibility handling is a reasonable and useful strategy. In particular, backward–forward interval propagation incorporates future constraints into current feasible ranges, helping exclude decisions that could lead to future constraint violations.

Computational efficiency. The proposed framework achieves millisecond-level inference and near-zero constraint violations on the evaluated benchmarks.

Critical weaknesses
  1. Incomplete related work. The discussion of learning-based constraint handling is incomplete. Several relevant works have investigated feasibility prediction, proactive infeasibility prevention, and efficient solution repair, including PIP/PIP-D [1], CaR [2], and Bisection Projection [3]. Although these methods target different optimization settings, they share similar motivations concerning efficient feasibility handling. The authors should clarify how SMOP differs from these approaches and better position its contributions within the existing literature.

  2. Limited experimental generalizability. Although SMOP is presented as a general framework for multi-period resource allocation, the experiments are restricted to fixed-commitment economic dispatch. The three benchmarks mainly differ in the number of generators and use the same planning horizon. Additional evaluations on other resource-allocation problems, longer planning horizons, and different constraint tightness levels would strengthen the generalizability claims.

  3. Insufficient experimental analysis. The experimental evaluation lacks sufficient analysis of the contributions of individual components. For example, Table 5 shows that removing the bound-feasibility loss improves the optimality gap while maintaining feasibility, raising questions about its necessity. Moreover, the computational comparisons would benefit from a clearer explanation of the batching settings and solver configurations.

  4. Insufficient comparison with self-supervised baselines. Although SMOP is proposed as a self-supervised optimization framework, the experiments include only one self-supervised baseline (GRU + CVXPY), while most comparisons are against supervised methods. A broader comparison with existing self-supervised optimization approaches would provide stronger evidence of the proposed method's effectiveness. References

[1] Bi et al. Learning to Handle Complex Constraints for Vehicle Routing Problems. NeurIPS, 2024.

[2] Bi et al. Towards Efficient Constraint Handling in Neural Solvers for Routing Problems. ICLR, 2026.

[3] Liang and Chen. Efficient Bisection Projection to Ensure Neural-Network Solution Feasibility for Optimization over General Set. ICML, 2025.

Questions for authors
  1. How does SMOP differ from existing feasibility-handling methods?

  2. Can SMOP be evaluated on other multi-period resource-allocation problems or under different planning horizons and constraint tightness levels?

  3. Proposition 1 assumes that feasible intervals exist at every period. How does the method handle cases where an earlier decision makes future intervals infeasible?

  4. Could the authors include additional self-supervised optimization baselines beyond GRU + CVXPY to better demonstrate the effectiveness of SMOP compared with existing approaches?

  5. Why does removing the bound-feasibility loss improve the optimality gap in Table 5? What additional benefit does this component provide?

  6. Could the authors clarify the computational comparison settings, particularly whether inference times are measured with comparable batching strategies?

5 comments
AuthorsOpenReview ↗

Response to the weakness 1 and questions 1 from Reviewer jDHx

Weakness 1 & Question 1: Incomplete related work...

Response: Thank you for these relevant references. We expanded the discussion to distinguish preventative feasibility handling, iterative repair, feasible-set parameterization, and differentiable optimization layers and added a paragraph in Section 5 of related work. Briefly summarize it here: Feasibility-aware neural optimization combines proactive constraint handling with prediction repair. For routing, PIP [1] uses penalties and action masking, while CaR [2] jointly trains construction and refinement policies. These methods improve empirical feasibility, but their discrete operators do not directly transfer to continuous allocation. Bisection Projection (BP) [3] restores continuous feasibility by searching along a segment between a prediction and a valid interior reference point. Although both BP and our method use bisection, BP searches for a segment coefficient, whereas our layer determines the balance multiplier for Euclidean projection onto a nonempty box-and-balance feasible set. Our framework combines this projection with backward--forward interval propagation and demand-aware bound tightening to support efficient recovery under temporal and system-wide coupling, without requiring a learned interior reference trajectory.

AuthorsOpenReview ↗

Response to the Weakness 2 and Question 2 from Reviewer jDHx

Weakness 2 & Questions 2: Limited experimental generalizability. Although SMOP is presented as a general framework for multi-period resource allocation, the experiments are restricted to fixed-commitment economic dispatch. The three benchmarks mainly differ in the number of generators and use the same planning horizon. Additional evaluations on other resource-allocation problems, longer planning horizons, and different constraint tightness levels would strengthen the generalizability claims.

Response: Thank you for this suggestion. We have added two benchmarks—supply-chain optimization and large-scale data-center workload allocation—and describe their formulations and methods in Sections 2 and 6 and Appendices F.4 and F.5. For the supply-chain benchmark, we introduce two variants, SMOP-HC and SMOP-HC-α\alpha, and analyze their trade-offs in objective gap, constraint violation, and inference time though both find feasible solution efficiently. For the large-scale data-center benchmark of 1310 units, we implement fast inference and report its speedup relative to a linear programming solver when the batch size is 1 or 8. We also include comprehensive ablation studies for both benchmarks in the appendix. These benchmarks demonstrate SMOP’s applicability across diverse multi-period optimization problems.

MethodMean abs. gap (pp)Max. violationTime (ms/sample)Speedup
SMOP-HC-α, float640.117\mathbf{0.117}<10−9<10^{-9}0.5171.47×1.47\times
SMOP-HC, float640.1220.122<10−9<10^{-9}0.505\mathbf{0.505}1.50×\mathbf{1.50\times}
ConfigurationMean costMean absolute gapMax. ViolationInference time (ms/instance)Speedup
SMOP, batch 120.630.0063$ < 10^{-9}$63.67±5.9463.67\pm5.9416.66×16.66\times
SMOP, batch 820.630.0063$ < 10^{-9}$34.16±1.0934.16\pm1.0931.06×31.06\times
HiGHS Solver20.5000$1060.9 $--
AuthorsOpenReview ↗

Response to the weakness 3 and question 6 from Reviewer jDHx

Weakness 3 & Question 6: Insufficient experimental analysis. The experimental evaluation lacks sufficient analysis of the contrib utions of individual components. For example, Table 5 shows that removing the bound-feasibility loss improves the optimality gap while maintaining feasibility, raising questions about its necessity. Moreover, the computational comparisons would benefit from a clearer explanation of the batching settings and solver configurations.

Response: Thanks for your suggestions. (1) we have supplemented more analysis on each component following the updated Table 5 in Section 6 and quoted as follows. Note that this Table 5 is updated as we applied unrestricted neural network predictor rather than the previous softmax activation function described in Section 3.2 highlighted in blue: We evaluate seven SMOP configurations to assess the contributions of individual components and loss weights to feasibility and solution quality. All configurations share the same dataset, GRU architecture, and label-free training framework. Figure 7 and Table 5 show that bound-feasibility regularization and sequential recovery are particularly important for controlling constraint violations. Removing either component substantially increases the maximum violation. The bound-feasibility loss encourages the aggregate reachability condition ∑gLg,tr≤Dt≤∑gUg,tr\sum_g L^r_{g,t} \leq D_t \leq \sum_g U^r_{g,t} used in Proposition 1. Sequential recovery incorporates ramping constraints by updating each period’s recovery bounds based on the preceding recovered dispatch. Removing backward propagation increases the objective gap, suggesting that tightening recovery intervals using future constraints helps guide the model toward better dispatch decisions. Feasibility-focused weights also produce a larger objective gap because they place greater relative emphasis on constraint penalties and less on operating cost. These results highlight the importance of balancing feasibility regularization with the cost objective: small constraint violations alone do not imply a high-quality solution.

(2) We have added the batching settings and solver configurations in Section 6 in blue. We have distinguished the single-instance inference time and batched averaged one-instance time in Section 6.2. Quote here: We compare the per-instance solution time of SMOP and the HiGHS linear-programming solver, showing speedups ranging from 11.39×11.39\times to 51.73×51.73\times. A key source of this acceleration is the batch-processing capability of neural networks, which allows many instances to be evaluated simultaneously and substantially reduces the amortized inference time per instance. Batch processing contributes substantially to SMOP’s computational advantage by evaluating multiple instances simultaneously and amortizing tensor-processing and framework overhead. For single-instance inference, SMOP requires 17.75±0.1217.75 \pm 0.12 , 17.78±0.1717.78 \pm 0.17 , and 18.06±0.6018.06 \pm 0.60 ms on OR-LIB 10, 50, and 100, respectively. Compared with the reported HiGHS runtimes, HiGHS remains faster on OR-LIB 10, whereas SMOP achieves lower single-instance latency on OR-LIB 50 and 100. For even larger data-center instances described in Section 6.4, SMOP also obviously outperforms the HiGHS LP solver under both batched and single-instance evaluation.

(3) Solver configuration. We apply the HiGHS through SciPy’s linprog interface (method="highs"), with presolve enabled and the remaining solver options left at their defaults. Machine activation, startup, and shutdown decisions are fixed before solving the continuous dispatch LP. We have added this description in Section 6.

ConfigurationMax. violation (MW)Abs. gap (%)
Full SMOP<10−9<10^{-9}0.039±2.31×10−30.039 \pm 2.31\times10^{-3}
No backward propagation<10−9<10^{-9}0.188±0.0240.188 \pm 0.024
No projection alignment<10−9<10^{-9}0.039±2.44×10−30.039 \pm 2.44\times10^{-3}
No bounds loss4.645±2.2154.645 \pm 2.2150.056±2.88×10−30.056 \pm 2.88\times10^{-3}
No sequential recovery6.187±0.3276.187 \pm 0.3270.069±0.0270.069 \pm 0.027
Cost-focused weights<10−9<10^{-9}0.021±1.04×10−30.021 \pm 1.04\times10^{-3}
Feasibility-focused weights<10−9<10^{-9}0.428±0.1950.428 \pm 0.195
Reviewer_hfa6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

SMOP is a self-supervised optimization proxy for multi-period resource allocation with balance, capacity, and ramping constraints. It combines backward–forward interval propagation, a BiGRU predictor, per-period balance projection, and sequential updates of reachable bounds. Reserve is allocated analytically from remaining headroom. Experiments on fixed-commitment economic dispatch report small objective gaps and constraint residuals, with faster batched inference than the evaluated optimization baselines.

Critical strengths
  1. The method makes useful use of the problem structure. Separating within-period balance repair from temporal constraints offers a practical alternative to a coupled projection. Analytic reserve allocation is also appropriate for the headroom-only reserve model considered.
  2. The reported feasibility and solution quality are encouraging. Table 2 shows essentially zero measured violations and reported gaps below 0.5% across three system sizes. On the 10-unit benchmark, SMOP improves both inference time and objective quality over the self-supervised CVXPYLayers baseline.
  3. The ablations examine relevant implementation choices. The studies cover backward propagation, sequential recovery, loss terms, numerical precision, and bisection iterations. Table 5 supports the importance of the temporal components, and the precision study helps distinguish round-off errors from larger violations.
Critical weaknesses
  1. The full-horizon guarantee depends on a condition the method does not establish. Proposition 1 assumes that every sequentially updated interval contains a demand-balanced dispatch. Componentwise bounds do not ensure that the current choice leaves a feasible continuation. When demand becomes unreachable, Algorithm 4 clips it, violating the original balance requirement. The theorem is valid conditionally, but does not establish that inference will satisfy its premise.
  2. The softmax output restricts which feasible solutions can be represented. With three generators bounded by [0,1][0,1] and demand 1.51.5, the raw outputs sum to one. Balance projection then shifts all components upward, leaving each at least 1/61/6. Thus the model cannot represent a feasible optimum such as (1,0.5,0)(1,0.5,0). This restriction comes from requiring the interval fractions to sum to one.
  3. The ramp formulas and pseudocode are inconsistent. Section 3.1 propagates backward using Lg,t+1−Φg,t+1L_{g,t+1}-\Phi_{g,t+1} and Ug,t+1+Ψg,t+1U_{g,t+1}+\Psi_{g,t+1}, but Algorithm 2 swaps the ramp directions and uses index tt. Algorithm 4 also uses index t−1t-1 for the transition into period tt. These differences affect the feasible set when ramp limits are asymmetric or vary over time.
  4. The speedup figures need to distinguish throughput from latency. Table 2 reports 0.80 ms per sample for the 10-unit system, while Figure 6 shows single-instance inference above 10 ms. The paper explains that batching contributes to the speedup, but these measurements do not establish the same advantage for an individual online decision.
Questions for authors
  1. What ensures that a balanced dispatch at the current period leaves a feasible continuation? If this is not guaranteed, could you state the additional assumption or fallback needed to avoid clipping demand at inference?
  2. Can the stated softmax-and-projection parameterization represent (1,0.5,0)(1,0.5,0) in the example above? If not, why impose a simplex constraint on the interval fractions rather than predict them independently?
  3. Which ramp directions and time indices does the implementation use? Please reconcile Algorithms 2 and 4 with the main equations.
  4. What are the batch-size-one latency and batched throughput under the same settings, including any instance-dependent bound propagation?
1 comment
AuthorsOpenReview ↗

Response to Reviewer hfa6

Weakness 1 & Question 1: The full-horizon guarantee ...

Response: We agree that the feasibility guarantee is conditional on the assumptions and bound conditions in Proposition 1. We clarify this limitation and introduce three complementary mechanisms to mitigate infeasibility.

(1) Next-period feasibility check and local correction. We have added Corollary 1 in Section 4, which introduces an explicit next-period feasibility check and a local repair. We first adjust the neural dispatch proposal to obtain pt(0)p_t^{(0)}. SMOP then computes the next-period bounds induced by $ p_t^{(0)} $, i.e., Lg,t+1r(pt(0))L^r_{g,t+1}(p_t^{(0)}) and Lg,t+1r(pt(0))L^r_{g,t+1}(p_t^{(0)}). It accepts this tentative dispatch only if every next-period interval is nonempty and ∑gLg,t+1r(pt(0))≤Dt+1≤∑gUg,t+1r(pt(0))\sum_g L^r_{g,t+1}(p_t^{(0)})\le D_{t+1}\le\sum_g U^r_{g,t+1}(p_t^{(0)}). If either condition fails, SMOP seeks a replacement dispatch by solving min⁡x\min_x $ ||x-\tilde{p_t}||^2$ s.t. ∑gxg=Dt\sum_g x_g=D_t, Lg,tr≤xg≤Ug,trL^r_{g,t}\leq x_g\leq U^r_{g,t}, Lg,t+1r(x)≤Ug,t+1r(x)L^r_{g,t+1}(x)\le U^r_{g,t+1}(x), ∑gLg,t+1r(pt(0))≤Dt+1≤∑gUg,t+1r(pt(0))\sum_g L^r_{g,t+1}(p_t^{(0)})\le D_{t+1}\le\sum_g U^r_{g,t+1}(p_t^{(0)}). Under the assumptions of Proposition 1, the final two constraints ensure that a next-period dispatch satisfy all the constraints. Once the bounds are available, the check requires O(G)O(G) operations; the local optimization is invoked only when the check fails.

(2) Full-horizon fallback. One-step extendability does not, by itself, establish full-horizon feasibility. If the local correction problem is infeasible, SMOP falls back to the original full-horizon LP, allowing earlier tentative dispatches to be revised. Provided that this LP is feasible and the solver returns a solution that passes the stated constraint checks, the fallback supplies a feasible full-horizon trajectory.

(3) Training regularization. The bound-feasibility loss, Lbounds\mathcal L_{\mathrm{bounds}}, complements these safeguards by penalizing demand outside the dynamically reachable interval. Because future bounds depend on earlier dispatches, this loss provides gradients that encourage decisions with feasible continuations, as explained in Proposition 4. This regularization can reduce the need for correction in training.

Our updated experiments provide empirical evidence under the evaluated settings. For example, all 301 DC-Full validation instances passed the independent constraint audit at a tolerance of 10−610^{-6}, with a maximum violation across all audited constraints of 5.60×10−75.60\times10^{-7}cores. These results support the effectiveness of the evaluated recovery pipeline without establishing an unconditional feasibility guarantee.

Weakness 2 & Questions 2: The softmax output ...

Response: We agree that requiring the interval fractions to sum to one introduces an unnecessary restriction. We adopted an unrestricted predictor in Section 3.2 in the revised paper: "The neural network predicts an unrestricted coordinate ηg,t\eta_{g,t} without softmax activation function, which is mapped to a dispatch proposal using the recovery bounds: p~g,t=Lg,tr+ηg,t(Ug,tr−Lg,tr).\tilde p_{g,t}=L^r_{g,t}+\eta_{g,t}\left(U^r_{g,t}-L^r_{g,t}\right).". Accordingly, we renewed all the experimental results related and indeed observed the improvement in tighter optimality gap and smaller maximum violation. Check Table 1 & 2.

Weakness 3 and Question 3: The ramp formulas and pseudocode ...

Response: We have corrected the writing to be consistent with the code. We can confirm that the code used the same ramping direction and transition from t-1 to t.

Weakness 4 & question 4: The speedup figures need to distinguish t...

Response: Yes, we added the single-instance inference time to distinguish it from batched one-instance inference time in Section 6.2 (highlight in blue). Batch processing contributes substantially to SMOP’s computational advantage by evaluating multiple instances simultaneously and amortizing tensor-processing and framework overhead. For single-instance inference, SMOP requires 17.75±0.1217.75 \pm 0.12, 17.78±0.1717.78 \pm 0.17, and 18.06±0.6018.06 \pm 0.60 ms on OR-LIB 10, 50, and 100, respectively. Compared with the reported HiGHS runtimes, HiGHS solver remains faster on small dataset OR-LIB 10, whereas SMOP achieves lower single-instance inference time on larger datasets of OR-LIB 50 and 100. For even larger data-center instances described in Section 6.4, SMOP also obviously outperforms the HiGHS LP solver under both batched and single-instance evaluation.

DatasetMax. violation (MW)Absolute gap (%)Train (s)Infer. (ms/sample)HiGHS total (ms/sample)Time ratio
OR-LIB 10<10−9<10^{-9}0.0283±0.004840.0283\pm0.0048471.8±0.471.8\pm0.40.683±0.0080.683\pm0.0087.78±0.2347.78\pm0.23411.39×
OR-LIB 50<10−9<10^{-9}0.1114±0.00720.1114\pm0.0072184.1±7.9184.1\pm7.91.218±0.0551.218\pm0.05538.92±3.6738.92\pm3.6731.94×
OR-LIB 100<10−9<10^{-9}0.1869±0.02290.1869\pm0.0229331.3±2.0331.3\pm2.01.567±0.0221.567\pm0.02281.08±7.8581.08\pm7.8551.73×

Comments on the paper

1 comment
AuthorsOpenReview ↗

Main improvement

We thank all reviewers for their thoughtful and constructive feedback. We have thoroughly revised the paper (changes highlighted in blue) and summarize the main updates below:

  1. Key improvement 1: Expanded experiments and scalability evaluation. We extend SMOP to two additional benchmarks: supply-chain optimization and large-scale data-center workload allocation. Sections 2 and 6, together with the appendix, present their formulations, implementation details, and comparisons with conventional optimization solvers. The data-center benchmark includes 1,310 units and 17,032 input features, demonstrating SMOP’s ability to handle high-dimensional instances.

  2. Key improvement 2: Unrestricted neural predictions and improved performance. We replace the softmax parameterization with an unrestricted neural prediction for each unit and period: p~g,t=Lg,tr+ηg,t(Ug,tr−Lg,tr)\tilde p_{g,t} =L^r_{g,t}+\eta_{g,t} (U^r_{g,t}-L^r_{g,t}), where ηg,t\eta_{g,t} is the output of neural networks without the softmax activation function. This affine mapping expresses the prediction in dispatch units without restricting the proposal to the recovery interval. The updated experiments in Section 6 show smaller optimality gaps and lower maximum constraint violations.

  3. Key improvement 3: Comparison with an additional self-supervised baseline. We implement DC3, a self-supervised method with constraint completion and correction, and compare it with SMOP in Section 6.1 and Table 1. Under the evaluated settings, SMOP achieves a higher feasibility rate, lower maximum constraint violations, smaller optimality gaps, and shorter inference times. The substantial inequality violations observed for DC3 highlight the difficulty of handling coupled constraints across multiple periods.

  4. Key improvement 4: Infeasibility analysis and a mitigation strategy. We analyze cases in which the feasibility conditions of Proposition 1 are not satisfied and introduce a local feasibility repair in Corollary 1. This repair serves as a backup, invoked only when an infeasible recovery step occurs. We prove that it resolves the specific counterexamples analyzed, clarifying both the limitations of the original recovery procedure and the scope of the proposed mitigation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.