Towards Reliable Neural Optimizers: Permutation-Equivariant Neural Approximation in Dynamic Data Driven Applications Systems

Javad Mohammadi; Meiyi Li

REVIEW 4 major objections 4 minor 16 references

Towards Reliable Neural Optimizers: Permutation-Equivariant Neural Approximation in Dynamic Data Driven Applications Systems

T0 review · 4 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A feed-forward neural optimizer with a permutation-equivariant attention module and gauge-map rescaling can produce feasible, near-optimal dispatch decisions in under a millisecond, regardless of sensor order.

desk verdict The permutation-equivariance proof is clean but standard, and the paper's central feasibility guarantee is false as written because Eq. (12) is not a gauge map. read the letter →

arxiv 2508.19364 v1 pith:OU6HRO5I submitted 2025-08-26 eess.SY cs.SY

Meiyi Li , Javad Mohammadi This is my paper · ORCID

classification eess.SYcs.SY

keywords learningtooptimizepermutationequivariancesensornetworksgaugemapfeasibilityguaranteevirtualpowerplantdynamicdatadrivenapplicationssystemsneuraloptimization

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

This paper argues that a feed-forward neural network can act as a real-time optimizer for dynamic sensor networks while respecting hard operational constraints. The proposed model, LOOP-PE, pairs a permutation-equivariant attention module with a generalized gauge map that rescales any candidate decision onto the feasible region. The central claim is a theorem: for any permutation of sensor inputs, the dispatch outputs are permuted consistently, and feasibility is preserved by construction. The paper supports this with a Virtual Power Plant case study in which LOOP-PE reaches decisions in about 0.33 ms on average, close to the baseline solver's optimum, with zero reported constraint violation.

What carries the argument

The load-bearing mechanism is the composition of two modules. The optimality module $O$ applies a shared fully connected embedding to each sensor, then a self-attention layer with shared $W_Q, W_K, W_V$ and no positional encodings; because row-wise softmax satisfies $\mathrm{softmax}(P_\sigma Z P_\sigma^\top) = P_\sigma \mathrm{softmax}(Z) P_\sigma^\top$, the whole module is permutation-equivariant. The feasibility module $T$ is the generalized gauge map $T(v) = u_0(x) + c(X, V) v$, where $c(X,V) = \left[\max_r \{1, [\sum_i H(x_i) v_i / \sum_i h(x_i)]_r\}\right]^{-1}$. That rescaling pushes any virtual prediction outside the feasible region onto the boundary of the constraint set while leaving feasible predictions unchanged. The paper assumes that $u_0(x)$ is an interior feasible, permutation-equivariant point; the gauge map then guarantees that every output satisfies the constraints in Eq. (2).

What would settle it

Take a two-sensor instance of the reformulated constraints in Eq. (2), choose an interior point $u_0(x)$ that satisfies the equalities, feed the network a virtual prediction far outside the feasible set, and check whether Eq. (12) maps it to a point satisfying every row of $\sum_i H(x_i) u_i \le \sum_i h(x_i)$; one instance where a violated row remains would directly refute the feasibility guarantee.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a neural optimizer can be built to be both permutation-equivariant and feasibility-preserving, end to end. Theorem 1 states that the full mapping $F = T \circ O$ satisfies $F(P_\sigma X) = P_\sigma F(X)$ for every permutation $\sigma$, where $O$ is the optimality module and $T$ is the gauge-based feasibility module. The proof rests on three facts: the per-sensor embedding commutes with permutations, self-attention without positional encodings is permutation-equivariant because row-wise softmax respects permutation matrices, and the gauge map's scaling factor is invariant under permutation while the interior point $u_0(x)$ is assumed equivariant. Consequently, the paper claims every output satisfies the local and coupled constraints of the reformulated linear optimization problem, so the neural dispatch can be implemented directly without iterative correction or post-processing.

Load-bearing premise

Everything rests on the assumption that a permutation-equivariant interior feasible point $u_0(x)$ exists, is available in closed form, and leaves enough slack that the gauge rescaling in Eq. (12) lands inside—not outside—the feasible set.

Editorial extensions

If this is right

If the theorem holds, no retraining or input pre-alignment is needed when sensors are reordered, dropped, or added; the output order automatically follows the input order.
Every dispatch decision satisfies local and coupled constraints by construction, so the neural optimizer can be used in applications where constraint violations are unacceptable.
Inference time can stay below one millisecond even when the baseline solver's runtime fluctuates, making the approach viable for real-time Dynamic Data Driven Applications Systems.
The architecture can handle a variable number of sensors because the constraints depend only on sums over sensors and the attention mechanism operates on sets rather than fixed-length vectors.

Reading between the lines

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

The proof of permutation equivariance does not depend on the specific attention weights, so the same guarantee should transfer to any set-equivariant encoder layer as long as the gauge map's assumptions are preserved.
Because the gauge map rescales linearly constrained regions, extending this approach to nonlinear or nonconvex constraints would require a new feasibility construction; that is the natural next test of the framework.
The reported zero feasibility gap covers the simulated VPP instances only; a stress test with adversarial or badly scaled inputs would show whether the gauge-map rescaling remains stable exactly at the constraint boundary.
If training happens to produce only feasible virtual predictions, the gauge rescaling never activates, so practical equivariance would reduce to the attention module alone; monitoring how often the rescaling fires would separate the two contributions.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Dear colleague,

The one thing to know: the permutation-equivariance proof here is correct but mostly known, while the paper's stronger claim—that the gauge map in Eq. (12) always produces feasible outputs—is wrong as written. The stress-test note is right, and the reader's conditional verdict is if anything too generous.

Credit where due. Theorem 1 is stated with explicit assumptions, and the proof that a pointwise feature embedding plus self-attention without positional encodings is permutation-equivariant is clean and easy to follow. This is a useful reference result, even if it is not new (refs [9,10,16] already cover it) and the paper admits it extends the authors' own earlier work. The virtual power plant application is sensible and the speed advantage of a feed-forward net over Gurobi is plausible.

The problem is Eq. (12), which defines the feasibility module as u = u0(x) + c v, with c computed from H v / h only. That is not a gauge map: it adds u0 to a scaled v instead of scaling the displacement v - u0. The counterexample from the stress-test note lands. One generator with u ≤ 1, H=1, h=1, u0=0.5, and a feasible virtual prediction v=0.8 gives Hv/h=0.8, so c=1, and the output is u=1.3 > 1. The slack already consumed by u0 never enters the formula. A correct radial projection would be u = u0 + c(v - u0) with c based on (Hv - Hu0)/(h - Hu0). As displayed, Eq. (12) does not guarantee feasibility, and the zero feasibility gap in Table 2 is unsupported.

Other soft spots amplify this. u0(X) is never defined or constructed; the paper just asserts it exists and is permutation-equivariant, then defers to preprint [13]. The ratio in Eq. (12) divides by sum h, which is zero for any lower-bound row (e.g., -u_i ≤ 0), so the formula is not even defined for the non-negativity constraints in the VPP case unless equalities are eliminated differently than the text suggests. The experiments add no training details, no architecture hyperparameters, no code or data; only 100 test samples, and optimality is measured against the same solver that produced the training labels.

So: the equivariance part is fine; the feasibility part is broken on the page. This paper is for readers who want a concise statement of permutation equivariance for attention-based optimizers. It is not for anyone who needs a working feasibility guarantee. I would send it back for major revision, asking the authors to correct Eq. (12) to a true gauge map, give a closed-form u0 for the VPP case, and release code and data. If they fix those, it becomes a reasonable application note. As submitted, the central claim is not supported.

Best, [You]

Referee Report

4 major / 4 minor

Summary. The paper presents LOOP-PE, a feed-forward, permutation-equivariant neural approximator intended to replace iterative optimization in real-time DDDAS settings. The architecture combines a shared pointwise embedding and self-attention without positional encodings (optimality module) with a feasibility module based on a 'generalized gauge map'. The formal contribution is Theorem 1, which states that the composition of these modules is permutation-equivariant. The experimental portion applies the method to a 20-DER virtual power plant dispatch problem, reporting sub-millisecond runtimes, a 4% average optimality gap, and a zero feasibility gap.

Significance. The equivariance proof in Section 4 is sound under the explicit assumptions, and the order-robustness and speed advantages are clearly demonstrated by the experiments. If the feasibility module worked as advertised, this would be a practically useful non-iterative optimizer for settings where constraint satisfaction is mandatory. However, the central feasibility guarantee is not established: Eq. (12) as written is not a valid feasibility-preserving gauge map, and the zero feasibility gap in Table 2 is not supported by the stated construction. The contribution is therefore conditional on repairing the feasibility module and re-validating the feasibility claim.

major comments (4)

[§3.2, Eq. (12); §4.2] The mapping in Eq. (12), u = u0(x) + c v with c = 1/max_r{1, [sum_i H(x_i)v_i / sum_i h(x_i)]_r}, is not a gauge map and does not guarantee feasibility. Consider a single constraint u <= 1 in the reformulated set (2), so H=1 and h=1. Let u0=0.5 and take a feasible virtual prediction v=0.8. Since Hv/h=0.8 <= 1, c=1 and the output is u=1.3, violating u <= 1. A radial projection from an interior point must scale the displacement v-u0, with a factor involving the residual slack h - sum_i H_i u0_i, and it should return v itself when v is already feasible. As written, Eq. (12) adds u0 to a scaled v instead of interpolating between u0 and v, so the statement in §3.2 that feasible virtual predictions are kept as they are is false, and the guaranteed-feasibility claim used for Table 2 and repeated in §6 is unsupported.
[Eq. (12); §2.1] The elementwise ratio in Eq. (12) divides by sum_i h(x_i), but the reformulated constraints in Eq. (2) can contain zero right-hand-side entries: for example, the lower bound 0 <= u_i is written as -u_i <= 0, giving h_i=0 for that row. Even in the VPP case study in Eq. (24b), the lower bounds 0 <= P_i^G introduce zero rows into the stacked h vector. The manuscript does not state the assumption h_i > 0 or define a convention for zero rows, so the formula is undefined for a natural class of problems, including the paper's own illustrative problem.
[§4.2, Theorem 1] The interior feasible point u0(x) is never defined or constructed; the manuscript only asserts that it exists, is interior, and is permutation-equivariant. This is a load-bearing assumption, not a harmless regularity condition: the feasibility of the final output depends on the slack provided by u0(x), and any corrected radial projection would require explicit knowledge of that slack. The authors should provide a closed-form construction of u0 for the problem class in (1), or state precisely for which constraint classes such an interior point exists and how it is computed, and they should prove its permutation equivariance rather than assuming it.
[§5.2, Table 2] Table 2 reports only the minimum feasibility gap (0.00) and lists no average or maximum, yet the text claims the feasibility gap remains consistently at 0.00 and the conclusion treats zero feasibility as established. Since Section 5.1 does not describe how constraint violations were measured or over how many test instances, the empirical zero-feasibility claim is not supported by the displayed data. The authors should report the full distribution of feasibility violations and the procedure used to check them.

minor comments (4)

[Eq. (12); §4.2] The notation in Eq. (12) writes T as a function of v only, but the formula depends on x through u0(x), H(x_i), and h(x_i); the notation should be T(v; x) or T(x, v), and the domain of T should be stated explicitly.
[Theorem 1; §2.2] Theorem 1 is stated only for a fixed number of sensors n and permutations in S_n, while the paper claims robustness to sensor dropout and system scaling, which involve changing n. A formal statement or argument covering variable n is needed to support those claims.
[§5.2, Eq. (25)] The optimality gap in Eq. (25) divides by ||u*||_2, which can be arbitrarily close to zero; the paper should report the range of u* or use an absolute gap.
[§3.2] The derivation of the gauge map in Eq. (12) is deferred to reference [13]; since the paper's advertised contribution is the analysis of the feasibility guarantee, the derivation (or a corrected version) should be reproduced here rather than deferred.

Circularity Check

2 steps flagged · score 6.0 of 10

Feasibility guarantee rests on self-cited prior work and the zero feasibility gap is tautological; the permutation-equivariance proof itself is self-contained.

self citation load bearing [Section 3.2, Eq. (12); Theorem 1 assumption (iii)]
"The point u0(x) is an interior feasible solution, which is itself permutation equivariant in x. Therefore, T defines a closed-form, permutation-equivariant feasibility mapping. A more detailed derivation is available in [13]."

Eq. (12) is the only mechanism offered for the paper's central feasibility guarantee, yet the formula's validity and the construction of the interior feasible point u0(x) are not derived here. The text simply asserts that u0(x) is an interior feasible solution and defers to [13], the authors' own prior work. Theorem 1 then assumes u0(X) is permutation-equivariant as an hypothesis rather than proving or constructing it. Thus the load-bearing feasibility claim is imported from a self-citation whose derivation is not reproduced, making the guarantee dependent on an unverified self-referential chain.
self definitional [Section 5.2, Table 2 and following paragraph]
"Furthermore, the feasibility gap remains consistently at 0.00, confirming that the proposed method strictly adheres to operational constraints."

Because the feasibility module T is defined as a gauge map that rescales outputs into the feasible set, a zero feasibility gap is entailed by the construction whenever the map is correct. Reporting this zero gap as 'confirming' strict constraint adherence treats the definition of the module as empirical evidence for its own correctness. The metric is not an independent test of feasibility; it can only detect whether the implementation matches the definition, not whether the feasibility guarantee is actually valid.

full rationale

The permutation-equivariance derivation in Sections 4.1-4.3 is self-contained and not circular: it proves pointwise-embedding equivariance and self-attention equivariance directly from the stated assumptions of shared weights and no positional encodings, and the composition argument T∘O is valid. That portion does not assume its conclusion. The circularity is confined to the feasibility side. Eq. (12) is presented as a generalized gauge map that guarantees feasible outputs, but the only support offered is a pointer to the authors' prior work [13], and the interior feasible point u0(x) is assumed to exist and be permutation-equivariant rather than constructed or proven. Consequently, the central advertised guarantee that LOOP-PE outputs satisfy constraints (1c)-(1d) is not independently established in this paper; it is inherited from a self-citation. The empirical zero feasibility gap in Table 2 is likewise an artifact of the construction rather than confirming evidence, since a correct gauge map makes feasibility true by definition. The optimality gap, by contrast, is a genuine benchmark against GUROBI and is not circular. Overall, the equivariance claim is independent, but the feasibility guarantee and its reported validation reduce to definition and to self-cited prior work, giving partial circularity. Separately, Eq. (12) as written appears mathematically questionable because it adds u0 to a scaled virtual prediction rather than interpolating toward u0, which is a correctness risk that reinforces the need for an actual derivation rather than a self-citation.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

No new physical entities are introduced. The ledger is dominated by the unstated construction of u0 and the deferred derivation of the gauge map, both inherited from prior work by the same authors. The neural network weights are fitted to Gurobi-generated labels, and the 10% fluctuation level is a hand-chosen data generation parameter.

free parameters (3)

neural network weights Theta = not reported
The optimality module is trained on Gurobi-generated samples; the learned weights determine solution quality and are fitted to data.
interior feasible point u0(x) = unspecified
Theorem 1 assumes u0 is permutation-equivariant and the gauge map depends on it, but the paper does not specify how this point is chosen or computed for a given sensor set.
data fluctuation level = 10%
The 10% fluctuation applied to capacity and demand is chosen by the authors, not derived from a physical dataset.

assumptions (6)

standard math Row-wise softmax is permutation-equivariant under simultaneous row and column permutation: softmax(P Z P^T) = P softmax(Z) P^T.
Used in Section 4.1(b) to show the attention output is equivariant.
standard math Self-attention with shared Q/K/V weights and no positional encodings is permutation-equivariant.
Standard property, proven in Section 4.1; assumed as a design choice.
domain assumption The original constraints (1c) and (1d) can be exactly reformulated as the sum-separable inequality (2) via variable elimination.
Section 2.1 states this without proof; the feasibility module's gauge map operates on this reformulation.
ad hoc to paper An interior feasible solution u0(x) exists and is permutation-equivariant.
Section 4.2 and Theorem 1 assume this; no construction or proof is provided in this manuscript.
ad hoc to paper The gauge-map scaling in Eq. (12) preserves feasibility of u0 + c v for all v.
The paper states this and refers to [13] for the derivation, but does not show it in this text; the displayed formula appears to ignore the slack consumed by u0.
domain assumption Training samples generated by adding 10% fluctuation to parameters from [14] and random sensor subsets are representative of deployment conditions.
Used in Section 5.1 to generate the 400-sample dataset; no real sensor data or out-of-distribution tests are used.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards Reliable Neural Optimizers: Permutation-Equivariant Neural Approximation in Dynamic Data Driven Applications Systems." pith.science (2026). https://pith.science/paper/OU6HRO5I

@misc{pith2026250819364,
  author       = {Pith},
  title        = {Pith review of: Towards Reliable Neural Optimizers: Permutation-Equivariant Neural Approximation in Dynamic Data Driven Applications Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OU6HRO5I}},
  note         = {Machine review of arXiv:2508.19364}
}

read the original abstract

Dynamic Data Driven Applications Systems (DDDAS) motivate the development of optimization approaches capable of adapting to streaming, heterogeneous, and asynchronous data from sensor networks. Many established optimization solvers, such as branch-and-bound, gradient descent, and Newton-Raphson methods, rely on iterative algorithms whose step-by-step convergence makes them too slow for real-time, multi-sensor environments. In our recent work, we introduced LOOP-PE (Learning to Optimize the Optimization Process, Permutation Equivariance version), a feed-forward neural approximation model with an integrated feasibility recovery function. LOOP-PE processes inputs from a variable number of sensors in arbitrary order, making it robust to sensor dropout, communication delays, and system scaling. Its permutation-equivariant architecture ensures that reordering the input data reorders the corresponding dispatch decisions consistently, without retraining or pre-alignment. Feasibility is enforced via a generalized gauge map, guaranteeing that outputs satisfy physical and operational constraints. We illustrate the approach in a DDDAS-inspired case study of a Virtual Power Plant (VPP) managing multiple distributed generation agents (DERs) to maximize renewable utilization while respecting system limits. Results show that LOOP-PE produces near-optimal, feasible, and highly adaptable decisions under dynamic, unordered, and distributed sensing conditions, significantly outperforming iterative algorithm based solvers in both speed and flexibility. Here, we extend our earlier work by providing additional analysis and explanation of LOOP-PE design and operation, with particular emphasis on its feasibility guarantee and permutation equivariance feature.

Figures

Figures reproduced from arXiv: 2508.19364 by the authors.

**Figure 2.** Building blocks of the proposed LOOP − PE model. The Optimality Module uses an attention mechanism to process input from varying sensor numbers and generate virtual predictions. The Feasibility Module uses the gauge map [13] to convert these predictions into practical, constraint-compliant actions, ensuring flexibility and robustness across different sensor setups and dynamics. 3.1 Optimality Module The optimality … view at source ↗

**Figure 3.** Solution spectrum for various test samples using data from multiple sensors. [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 12 canonical work pages

[13]

Toward Rapid, Optimal, and Feasible Power Dispatch through Generalized Neural Mapping

Li, Meiyi, and Javad Mohammadi. "Toward Rapid, Optimal, and Feasible Power Dispatch through Generalized Neural Mapping." arXiv preprint arXiv:2311.04838 (2023)

arXiv 2023
[1]

Sensor/Information Fusion, and Target Recognition XXXI

Kadar, Ivan, Erik P. Blasch, and Lynne L. Grewe. "Sensor/Information Fusion, and Target Recognition XXXI." Proc. of SPIE Vol. Vol. 12122. 2020

work page 2020
[2]

Information fusion as an autonomy enabler for uas traffic management (utm)

Blasch, Erik, et al. "Information fusion as an autonomy enabler for uas traffic management (utm)." Proceedings of the AIAA Scitech Forum. 2021

work page 2021
[3]

ImpactofDataPoisoningAttacksonFeasibility and Optimality of Neural Power System Optimizers

Agah,Nora,MeiyiLi,andJavadMohammadi."ImpactofDataPoisoningAttacksonFeasibility and Optimality of Neural Power System Optimizers." 2025 IEEE Texas Power and Energy Conference (TPEC). IEEE, 2025

work page 2025
[4]

Data Poisoning: An Overlooked Threat to Power Grid Resilience

Agah, Nora, et al. "Data poisoning: an overlooked threat to power grid resilience." arXiv preprint arXiv:2407.14684 (2024)

work page Pith review arXiv 2024
[5]

Meta optimal transport

Amos, Brandon, et al. "Meta optimal transport." arXiv preprint arXiv:2206.05262 (2022)

arXiv 2022
[6]

Learning to solve optimization problems with hard linear constraints

Li, Meiyi, Soheil Kolouri, and Javad Mohammadi. "Learning to solve optimization problems with hard linear constraints." IEEE Access (2023)

work page 2023
[7]

Learningassistedagent-basedenergyoptimization:Areinforcementlearning based consensus+ innovations approach

Du,Yuhan,etal."Learningassistedagent-basedenergyoptimization:Areinforcementlearning based consensus+ innovations approach." 2022 North American Power Symposium (NAPS). IEEE, 2022

work page 2022

Show all 16 references

[8]

Machinelearning/artificialintelligenceforsensordatafusion–opportunities and challenges

Blasch,Erik,etal."Machinelearning/artificialintelligenceforsensordatafusion–opportunities and challenges." IEEE Aerospace and Electronic Systems Magazine 36.7 (2021): 80-93

2021
[9]

Permutation Equivariance of Transformers and Its Applications

Xu, Hengyuan, et al. "Permutation Equivariance of Transformers and Its Applications." Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024

2024
[10]

Gauge equivariant transformer

He, Lingshen, et al. "Gauge equivariant transformer." Advances in Neural Information Pro- cessing Systems 34 (2021): 27331-27343

2021
[11]

Teachingnetworkstosolveoptimizationproblems

Liu,Xinran,etal."Teachingnetworkstosolveoptimizationproblems."IEEEAccess(2024)

2024
[12]

DeepOPF+:AdeepneuralnetworkapproachforDCoptimalpowerflow for ensuring feasibility

Zhao,Tianyu,etal."DeepOPF+:AdeepneuralnetworkapproachforDCoptimalpowerflow for ensuring feasibility." 2020 IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids (SmartGridComm). IEEE, 2020

2020
[14]

Machine Learning Infused Distributed Optimization for Coordinating Virtual Power Plant Assets

Li, Meiyi, and Javad Mohammadi. "Machine Learning Infused Distributed Optimization for Coordinating Virtual Power Plant Assets." arXiv preprint arXiv:2310.17882 (2023)

2023 arXiv
[15]

Gurobi Optimizer Reference Manual,

Gurobi Optimization, LLC, “Gurobi Optimizer Reference Manual,” 2023

2023
[16]

Permutation equivariant neural functionals

Zhou, Allan, et al. "Permutation equivariant neural functionals." Advances in neural infor- mation processing systems 36 (2023): 24966-24992

2023

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [13]

Toward Rapid, Optimal, and Feasible Power Dispatch through Generalized Neural Mapping

Li, Meiyi, and Javad Mohammadi. "Toward Rapid, Optimal, and Feasible Power Dispatch through Generalized Neural Mapping." arXiv preprint arXiv:2311.04838 (2023)

arXiv 2023

[2] [1]

Sensor/Information Fusion, and Target Recognition XXXI

Kadar, Ivan, Erik P. Blasch, and Lynne L. Grewe. "Sensor/Information Fusion, and Target Recognition XXXI." Proc. of SPIE Vol. Vol. 12122. 2020

work page 2020

[3] [2]

Information fusion as an autonomy enabler for uas traffic management (utm)

Blasch, Erik, et al. "Information fusion as an autonomy enabler for uas traffic management (utm)." Proceedings of the AIAA Scitech Forum. 2021

work page 2021

[4] [3]

ImpactofDataPoisoningAttacksonFeasibility and Optimality of Neural Power System Optimizers

Agah,Nora,MeiyiLi,andJavadMohammadi."ImpactofDataPoisoningAttacksonFeasibility and Optimality of Neural Power System Optimizers." 2025 IEEE Texas Power and Energy Conference (TPEC). IEEE, 2025

work page 2025

[5] [4]

Data Poisoning: An Overlooked Threat to Power Grid Resilience

Agah, Nora, et al. "Data poisoning: an overlooked threat to power grid resilience." arXiv preprint arXiv:2407.14684 (2024)

work page Pith review arXiv 2024

[6] [5]

Meta optimal transport

Amos, Brandon, et al. "Meta optimal transport." arXiv preprint arXiv:2206.05262 (2022)

arXiv 2022

[7] [6]

Learning to solve optimization problems with hard linear constraints

Li, Meiyi, Soheil Kolouri, and Javad Mohammadi. "Learning to solve optimization problems with hard linear constraints." IEEE Access (2023)

work page 2023

[8] [7]

Learningassistedagent-basedenergyoptimization:Areinforcementlearning based consensus+ innovations approach

Du,Yuhan,etal."Learningassistedagent-basedenergyoptimization:Areinforcementlearning based consensus+ innovations approach." 2022 North American Power Symposium (NAPS). IEEE, 2022

work page 2022