Neural Event-Triggered Control with Optimal Scheduling

Jingdong Zhang; Luan Yang; Qunxi Zhu; Wei Lin

REVIEW 4 major objections 6 minor 1 cited by

Neural Event-Triggered Control with Optimal Scheduling

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a neural controller trained with an event-triggered mechanism and an analytical projection can stabilize nonlinear systems with a handful of control updates while guaranteeing exponential stability and maximal…

desk verdict Strong empirical package with an overstated stability guarantee; the projection theorem in Section 4 does not hold as written and is not connected to the trained controllers. read the letter →

arxiv 2507.14653 v1 pith:KMFCQCGN submitted 2025-07-19 math.OC

Luan Yang , Jingdong Zhang , Qunxi Zhu , Wei Lin This is my paper · ORCID

classification math.OC MSC 93C5793D1593D0568T07

keywords event-triggeredcontrolneuralLyapunovoptimalschedulingminimalinter-eventtimeprojectionoperatorpathintegralalgorithmMonteCarlotrainingresource-constrained

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

The paper tries to show that event-triggered controllers can be learned end-to-end so that a stabilizing neural controller updates the control signal only a few times, while preserving an exponential stability guarantee and near-optimal scheduling. The proposed Neural ETC framework comes with two training algorithms: a path-integral algorithm that differentiates through event times computed by a neural event ODE solver, and a Monte Carlo algorithm that regularizes Lipschitz constants to enlarge the lower bound on inter-event time. A projection operator with an analytical expression is introduced to force the learned controller into the stable set, which the paper claims guarantees stability and a positive minimal inter-event time. If the claimed guarantee holds, learning-based feedback could be deployed on digital, communication-limited platforms without continuously streaming control values, and would keep the low online cost of an offline policy while responding in real time through the event function.

What carries the argument

The framework's engine is the augmented closed-loop dynamics $\dot x=f(x,u(x+e)),\; \dot e=-f(x,u(x+e))$ with error state $e=x(t_k)-x(t)$, which turns the open-loop event-triggered system into a continuous system whose triggering times are roots of the event function $h$. The event function is the mechanism that balances decay of the Lyapunov function against the tolerated error, giving an exponential decay rate $1-\sigma$. The lower-bound estimate derived in Theorem 3.2 uses a differential inequality for the ratio $z=\|e\|/\|x\|$ to produce a logarithmic lower bound on inter-event time, and Theorem 3.3 adapts that bound to the stability-based event function using the Lipschitz constants of $u$ and of the inverse class-$K$ function $\alpha^{-1}$. The projection $\pi$ is the mechanism that turns soft-trained neural candidates into certified controllers: after projection, the Lyapunov inequality holds by construction in the affine-actuation case, and boundedness of the state space gives Lipschitz continuity. The two training algorithms push inter-event time in different ways: path integral back-propagates through the computed triggering times, while Monte Carlo penalizes the Lipschitz constants that appear in the lower-bound formula.

What would settle it

On a dense grid in the state domain $D$, compute the post-projection Lyapunov derivative $\nabla V(x)\cdot f(x,\pi(u)(x))+V(x)$ for the identity-actuated Lorenz experiment; the stability claim fails if any sampled point gives a positive value. For the state-dependent benchmarks, repeat with $f(x,g(x)\pi(u)(x))$; if violations appear there but not in the additive case, the gap lies in the affine-actuation assumption rather than in the training procedure.

Watch

Extended reading notes

Core claim

Working from the standard neural Lyapunov control setup, the paper's central object is an event function $h(x,e)=\nabla V(x)\cdot\big(f(x,u(x+e))-f(x,u(x))\big)-\sigma V(x)$, under which the controlled system satisfies $\nabla V(x)\cdot f(x,u(x+e))\leq -(1-\sigma)V(x)$, so $V$ decays exponentially with rate $1-\sigma$ between updates. The projection $\pi(u,U(V))=u-\frac{\max(0,L_{f_u}V-V)}{\|\nabla V\|^2}\nabla V$ maps any candidate controller into $U(V)=\{u:L_{f_u}V+V\leq 0\}$ when actuation is affine, and is Lipschitz on bounded domains; with this projection, the triggering mechanism yields exponential stability and a positive lower bound on inter-event time. The paper further claims that minimizing the Lipschitz constant of the projected controller is a necessary condition for the largest minimal inter-event time, i.e., schedule optimality. Empirically, on a two-node gene regulatory network, the Lorenz system, and a 100-dimensional Michaelis-Menten network, Neural ETC variants report a handful of triggers over the control horizon and minimal inter-event times far above baselines, with low mean-squared error under a limit of ten triggers.

Load-bearing premise

The stability certificate is proven for controllers that enter the dynamics additively, while the reported gene-regulation and cell benchmarks have state-dependent actuation terms, so the guarantee as written does not literally cover those experiments.

Editorial extensions

If this is right

Neural ETC can be implemented on platforms where the control value is updated only at a few state-dependent instants rather than continuously, drastically reducing communication cost.
The Monte Carlo variant avoids back-propagation through ODE solvers, so training scales better for high-dimensional systems such as the 100-dimensional cell model.
The projection operation gives a fast, analytical way to certify a learned controller, without solving an optimization problem after training.
Regularizing the Lipschitz constant of the controller, rather than the convexity of the Lyapunov function, is the training factor that most directly reduces the number of triggers.
Event-triggered control inherits the low online cost of an offline policy while retaining the responsiveness of an online solver through the event function.

Reading between the lines

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

Because the projection proof treats control as additive, extending Neural ETC to underactuated state-dependent actuation, as in the GRN and cell benchmarks, requires a projection formula built from $\nabla V\cdot g(x)u$; this extension is implicit in the experiments but not proved in the paper.
For stochastic differential equations, where the triggering time is a stopping time that varies by sample path, one could train a Neural ETC by estimating the distribution of $t_1$ instead of a single root; the paper identifies this as an open difficulty.
The lower-bound analysis suggests a transferable design rule: penalizing the slopes of the controller and of the inverse class-$K$ decay function is a cheap proxy for communication cost in any certificate-based training loop.
A converse question is left open: Theorem 4.2 gives a necessary condition for the largest minimal inter-event time, so a full optimality certificate would need a matching upper bound or a global optimizer over the stable controller set.

Share X Bluesky LinkedIn Reddit HN

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

The thing to know: this is a genuinely useful empirical paper whose theory section does not support the advertising. The Neural ETC framework—two algorithms, path-integral and Monte-Carlo, for learning event-triggered neural Lyapunov controllers—is a real contribution to the ETC literature. The experiments are extensive: GRN, Lorenz, 100-D Michaelis-Menten, with error bars over five runs, and the code is released. On the empirical side, the claim that you can stabilize a 100-D system with two control updates while keeping low MSE is noteworthy and, as far as I can tell, credibly supported.

The soft spot is Theorem 4.1. As printed, the projection formula π(u,U(V)) = u − max(0,L_fuV−V)/‖∇V‖²·∇V is not dimensionally valid for the actuation used in the benchmarks: in GRN and the Cell model, u is scalar or multiplies a state-dependent g(x), so subtracting a d-dimensional gradient from u makes no sense. The proof goes further and assumes f(x,u)=f(x)+u, which none of the test systems satisfy. There is also a sign slip: the theorem writes L_fuV−V inside the max, the proof uses L_fuV+V. With the printed sign, the simple system ẋ=x+u, V=x², u=0 gives π=−x/2 and L_πV=V, not ≤−V. Fixing the sign still leaves the actuator issue. And neither Algorithm 1 nor Algorithm 2 ever applies this projection, so even a correct theorem would be disconnected from the trained controllers.

Theorem 4.2 is asserted without a proof and does not follow from Theorem 3.2's lower bound; it identifies the Lipschitz-minimal controller in U(V), which is at best a proxy for actual trigger-count minimization. The MC algorithm's 'optimality' is therefore an indirect regularization objective, not a demonstrated property.

None of this kills the empirical contribution. The path-integral method with neural event ODE is a clever training device, and the lower-bound-based MC regularizer is a sensible way to encourage larger inter-event times. The ablation and the comparison across methods are informative. But the abstract's claim that the projection 'ensures theoretical stability and schedule optimality' is not true for what is implemented.

This paper deserves a serious referee, but the verdict should be major revision, not acceptance. The authors need to either fix the projection theorem so it covers the actual actuation structure (or restrict its scope), correct the sign error, and connect the projection to the training pipeline, or remove the theoretical guarantee claims and present the result as an empirical method with heuristic regularization. For the right reader—someone working in learning-based event-triggered control—this is worth a careful look. I would bring it to reading group and I'd likely cite the empirical comparison if I cite it at all.

Referee Report

4 major / 6 minor

Summary. The paper proposes Neural ETC, a framework for learning event-triggered stabilizing controllers with Lyapunov certificate functions, claiming to achieve minimal triggering times and maximal inter-event times. Two training algorithms are presented: Neural ETC-PI, based on differentiable ODE event solvers, and Neural ETC-MC, based on a theoretical lower bound on inter-event time. The paper further introduces a projection operator that is claimed to ensure theoretical stability and schedule optimality, and it reports experiments on a two-node gene regulatory network, the Lorenz system, and a 100-dimensional Michaelis-Menten cell model, with code released. The main advertised contribution is that the learned controller stabilizes the system with the least number of control updates while retaining an exponential stability guarantee.

Significance. Event-triggered control with learned certificates is a timely and important topic, and the empirical results are striking: on the GRN benchmark, Neural ETC-MC uses 4 triggers versus 1816 for LQR, and on the 100-dimensional cell model it uses 2 triggers versus 449. The release of code and the report of error bars over 5 runs are commendable and facilitate reproducibility. If the theoretical guarantees were valid, the paper would be a significant contribution to networked and resource-constrained control. However, the stability and optimality guarantees advertised in the title, abstract, and Section 4 are not supported by the proofs as written, and the reported experiments are not covered by the projection theorem. The empirical comparison may survive as an engineering contribution after a substantial revision, but the theoretical claims need major correction or removal.

major comments (4)

[Theorem 4.1; Appendix A.1.4] The projection theorem is not valid as printed and does not cover the experimental systems. The definition pi(u,U(V)) = u - max(0, L_fu V - V)/||grad V||^2 * grad V uses the gradient direction, which only makes sense when the control enters as f(x)+u, i.e., with identity actuation. Appendix A.1.4 confirms this: the proof writes L_fu V|_{u=pi} = grad V . (f + u - max(0, L_u V + V)/||grad V||^2 * grad V). The sign in the printed theorem and the sign in the proof are also inconsistent: the theorem has max(0, L_fu V - V) while the proof uses max(0, L_u V + V). With the printed sign, the simple system xdot = x + u, V = x^2, u = 0 yields pi = -x/2 and L_{f+pi} V = V, which is not <= -V as claimed. In the GRN benchmark (Appendix A.3.2) the scalar control u multiplies x_1^n/(s^n+x_1^n), and in the Cell benchmark (Appendix A.3.4) the control multiplies diag(x_i^2/(1+x_i^2)); for these systems the projection formula is either undefined or corrects the wrong control direction. Consequently, Theorem 4.1 cannot certify stability for the reported experiments.
[Theorem 4.2; Eqs. (13)-(14)] The optimality guarantee is asserted without proof and does not follow from the lower-bound estimate in Theorem 3.2. Theorem 3.2 provides a lower bound on the minimal inter-event time that decreases with the controller Lipschitz constant l_u, but maximizing a lower bound is not equivalent to maximizing the actual inter-event time or minimizing the number of triggering events in the problem (2). Equation (14) states that pi(u,U(V)) belongs to argmin l_{pi(u,U(V))}, which is tautological unless a minimization over U(V) is actually performed; no such minimization is proved. The training objectives are 1/t_1 in Algorithm 1 and Lipschitz penalties in Algorithm 2, so even the optimization criterion used in training is only a proxy for the stated goal of minimal triggering times.
[Algorithms 1-2; Section 5 and Appendix A.3] The projection operation is never applied in the algorithms or the experiments. Algorithm 1 returns u_phi and V_theta directly, and Algorithm 2 returns u_phi, V_theta, and alpha_{theta_alpha}; the test configurations in Appendices A.3.2 through A.3.4 use event functions based on the learned u_phi and V_theta without any projection step. Therefore, even if Theorem 4.1 were corrected for identity actuation, it would not establish stability of the reported Neural ETC trajectories. The abstract's claim that the projection 'ensures theoretical stability and schedule optimality for Neural ETC' is disconnected from the implementation described in the paper.
[Theorem 4.1, Lipschitz-continuity claim] The 'if and only if' characterization of Lipschitz continuity is false. If L_fu V <= -V holds on an unbounded domain D, then the max term in the projection vanishes, so pi(u,U(V)) = u, which is Lipschitz on D by the standing assumption on u. The proof's argument that sup max(0, L_u V + V)/||grad V|| is finite if and only if D is bounded ignores this case, and the asymptotic estimate L_u V + V approximately O(||x||^p) is not justified for general u and V. In addition, the claim of a positive lower bound on inter-event time is not proven: it is asserted to follow from Theorem 3.2, but the hypotheses of Theorem 3.2 (the Lipschitz conditions and the input-to-state inequality involving alpha and gamma) are not verified for the projected controller and for the V-based event function used in Theorem 4.1.

minor comments (6)

[Remark 2.2] The sentence beginning 'Unlike model-free reinforcement learning (RL) approaches that search...' is grammatically incomplete; the first clause needs to be integrated into a full sentence.
[Appendix A.3.1] The symbol sigma is used both for the exponential decay parameter in Eqs. (3) and (12) and for the smoothed ReLU activation in the network definitions; this dual use is confusing and should be resolved by renaming one of them.
[Algorithm 2] The dataset-generation line '{xi}^N times {xi}^{M_alpha}' is ambiguous: the two samples are independent draws from mu(D) and mu(X), so they should be written as two separate sampling operations.
[Table 1] The Neural ETC-MC row appears to merge multiple numerical entries into a single cell (e.g., '15.52...0.07...'); the formatting should be corrected so that each metric has a separate entry.
[Theorem 3.2] The displayed formula for tau_h in the main text should be checked against the proof in Appendix A.1.2, since the denominator appears in different forms in the two places and the reader has to reconcile the notation.
[Section 5.4] The quantity Tr(grad^2 V) is called the 'convexity of V', but the trace of the Hessian is not a valid measure of convexity; positivity of the trace does not imply convexity. A different diagnostic should be used or the terminology should be corrected.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: stability and inter-event-time bounds are derived from stated assumptions; the theorem-level gaps are correctness errors, not self-referential reductions.

full rationale

The paper's derivation chain is not circular. The stability guarantee uses the external Lyapunov theorem (Theorem 2.1) and the event function (3)-(4), both standard event-triggered-control constructions. The lower bound in Theorems 3.2-3.3 is proved in Appendices A.1.2-A.1.3 from explicit Lipschitz assumptions, not assumed from the conclusion. The projection operation in Theorem 4.1 is an algebraic construction aimed at membership in U(V); however, the printed formula uses max(0, L_fu V - V) while the proof in A.1.4 uses max(0, L_u V + V), and the proof writes the Lie derivative as ∇V·(f+u), i.e., identity actuation, although the theorem claims general affine actuators. These are correctness errors, not circular reductions. Theorem 4.2's optimality claim is asserted as a direct result with no proof, and Eq. (14) as printed is either vacuous or unproved; this is an unsupported inference that substitutes Lipschitz minimization for the true inter-event-time objective, but it is not a fitted parameter renamed as a prediction. Self-citations appear in the parametrization adopted from (Zhang et al., 2022a; 2024b) and in the domain-motivation paragraphs, but the parametrization is the standard ICNN construction and the cited prior work is not the load-bearing justification for the main result. The empirical comparisons are self-contained against external benchmarks, so the paper is not circular; score 2 reflects only minor non-load-bearing self-citations, while the theorem-level problems noted should be handled as correctness risks.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claims rest on standard Lyapunov stability theory, Lipschitz regularity of the vector field and controller, and the assumption that the neural networks trained with soft losses actually satisfy the certificate inequalities on the relevant domain. The handwritten hyperparameters (sigma, lambda2, epsilon) influence the results, and the optimality claim relies on the assumption that the theoretical lower bound is a faithful proxy for the true inter-event time.

free parameters (5)

sigma (exponential decay parameter) = 0.5 (default), 0.8 (ablation best)
Hand-tuned in event function; trades decay rate against inter-event time.
lambda2 (event loss weight) = 0.05 (PI), 0.1 (MC) in Table 1
Hand-tuned; large values improve scheduling but break stabilization (Table 3).
epsilon (quadratic regularization in V) = 1e-3
Default chosen to make V positive definite; used in all experiments.
lambda1 (Lipschitz regularization weight) = not tuned
Set by default in the spectral norm package; the paper states it does not tune it.
learning rates, batch sizes, iterations = listed per experiment in Appendix A.3
Hand-chosen per benchmark; affect the trained certificates.

assumptions (6)

standard math Lyapunov stability conditions (Theorem 2.1): existence of V with V(0)=0, V(x)>=c||x||^p, L_fu V <= -delta V.
Foundation for the stability guarantees; cited from Mao (2007).
domain assumption Vector field f and controller u are Lipschitz with constants lf and lu (conditions (i)-(ii) of Theorem 3.2).
Needed for the inter-event time lower bound; assumed for the systems under study.
domain assumption There exist class-K functions alpha and gamma satisfying L_fu V(x,u(x+e)) <= -alpha(||x||) + gamma(||e||) with alpha^{-1}(gamma(||e||)) <= P||e||.
Used to derive the lower bound; the paper enforces only a sufficient split (Eqs. 7-8) via soft losses.
domain assumption State space D is bounded (Theorem 3.3 and Lipschitzness of the projection).
Used to bound max ||grad V|| and to ensure the projected controller is Lipschitz.
ad hoc to paper The controller enters the dynamics affinely with identity actuation (f + u) in the projection proof.
The proof of Theorem 4.1 in Appendix A.1.4 writes Lie derivative as grad V . (f + u - ...); the test systems use state-dependent actuation g(x) not equal to identity.
ad hoc to paper The trained neural networks (Vtheta, u_phi, alpha_theta) trained with soft losses satisfy the certificate inequalities on the relevant domain.
The paper relies on this for stability, but only soft penalties are used and the projection is not applied in the algorithms.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Neural Event-Triggered Control with Optimal Scheduling." pith.science (2026). https://pith.science/paper/KMFCQCGN

@misc{pith2026250714653,
  author       = {Pith},
  title        = {Pith review of: Neural Event-Triggered Control with Optimal Scheduling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KMFCQCGN}},
  note         = {Machine review of arXiv:2507.14653}
}

read the original abstract

Learning-enabled controllers with stability certificate functions have demonstrated impressive empirical performance in addressing control problems in recent years. Nevertheless, directly deploying the neural controllers onto actual digital platforms requires impractically excessive communication resources due to a continuously updating demand from the closed-loop feedback controller. We introduce a framework aimed at learning the event-triggered controller (ETC) with optimal scheduling, i.e., minimal triggering times, to address this challenge in resource-constrained scenarios. Our proposed framework, denoted by Neural ETC, includes two practical algorithms: the path integral algorithm based on directly simulating the event-triggered dynamics, and the Monte Carlo algorithm derived from new theoretical results regarding lower bound of inter-event time. Furthermore, we propose a projection operation with an analytical expression that ensures theoretical stability and schedule optimality for Neural ETC. Compared to the conventional neural controllers, our empirical results show that the Neural ETC significantly reduces the required communication resources while enhancing the control performance in constrained communication resources scenarios.

Figures

Figures reproduced from arXiv: 2507.14653 by the authors.

**Figure 1.** Comparison of Neural ETC (yellow) and neural Lyapunov control (NLC, purple) in stabilizing the Lorenz system under the event-triggered control setting. (a): The inter-event time of consecutive triggering events. The dashed lines represent the minimal inter-event time of each control. (b): The control values acting on variable x in the control process. (c): The controlled trajectories of the variable x. engineering.… view at source ↗

**Figure 2.** Illustration of the Neural ETC with optimal scheduling. design the event function h = h(x, e) as h = ∇V (x) · (f(x,u(x + e)) − f(x,u(x))) − σV (x) (3) with 0 < σ < 1, such that the event-triggered controlled system satisfies ∇V (x) · f(x,u(x + e)) ≤ ∇V (x) · f(x,u(x)) + σV (x) ≤ −(1 − σ)V (x). (4) Hence, the exponential stability of the event-triggered controlled system is assured with exponential decay rate 1 − σ.… view at source ↗

**Figure 4.** (a) Convexity of V is calculated as the trace of the ∇2V on 1000 points in [−2.5, 2.5]3 . (b) Triggering times and (c) norm of ∇u in the training process. 6. Related Work Neural control with certificate functions. Previous works in neural control establish the performance guarantee via using the certificate functions, including Lyapunov function for stability (Giesl & Hafstein, 2015; Chang et al., 2019), barrier fu… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Lyapunov Guidance: A Unified Framework for Stabilizing Generative Flows
cs.LG 2026-07 reject novelty 5.0 of 10

Flow guidance is framed as Lyapunov control with a pseudo-projection for stability, but the projected flow is not shown to sample the target conditional distribution.

Reference graph

Works this paper leans on

57 extracted references · 54 canonical work pages · cited by 1 Pith paper

[1]

A simple event-based pid controller

arz \'e n, K.-E. A simple event-based pid controller. IFAC Proceedings Volumes, 32 0 (2): 0 8687--8692, 1999

work page 1999
[2]

Stabilization of nonlinear systems using event-triggered output feedback controllers

Abdelrahim, M., Postoyan, R., Daafouz, J., and Ne s i \'c , D. Stabilization of nonlinear systems using event-triggered output feedback controllers. IEEE transactions on automatic control, 61 0 (9): 0 2682--2687, 2015

work page 2015
[3]

D., Xu, X., Grizzle, J

Ames, A. D., Xu, X., Grizzle, J. W., and Tabuada, P. Control barrier function based quadratic programs for safety critical systems. IEEE Transactions on Automatic Control, 62 0 (8): 0 3861--3876, 2016

work page 2016
[4]

Amos, B., Xu, L., and Kolter, J. Z. Input convex neural networks. In International Conference on Machine Learning, pp.\ 146--155. PMLR, 2017

work page 2017
[5]

str \"o m, K. J. and Bernhardsson, B. Comparison of periodic and event based sampling for first-order stochastic systems. IFAC Proceedings Volumes, 32 0 (2): 0 5006--5011, 1999

work page 1999
[6]

Deep reinforcement learning for event-triggered control

Baumann, D., Zhu, J.-J., Martius, G., and Trimpe, S. Deep reinforcement learning for event-triggered control. In 2018 IEEE Conference on Decision and Control (CDC), pp.\ 943--950. IEEE, 2018

work page 2018
[7]

The control of chaos: theory and applications

Boccaletti, S., Grebogi, C., Lai, Y.-C., Mancini, H., and Maza, D. The control of chaos: theory and applications. Physics reports, 329 0 (3): 0 103--197, 2000

work page 2000
[8]

Neural lyapunov control

Chang, Y.-C., Roohi, N., and Gao, S. Neural lyapunov control. In Proceedings of the 33rd International Conference on Neural Information Processing Systems, pp.\ 3245--3254, 2019

work page 2019

Show all 57 references

[9]

T., Amos, B., and Nickel, M

Chen, R. T., Amos, B., and Nickel, M. Learning neural event functions for ordinary differential equations. In International Conference on Learning Representations, 2020

2020
[10]

Event-triggered optimal nonlinear systems control based on state observer and neural network

Cheng, S., Li, H., Guo, Y., Pan, T., and Fan, Y. Event-triggered optimal nonlinear systems control based on state observer and neural network. Journal of Systems Science and Complexity, 36 0 (1): 0 222--238, 2023

2023
[11]

and Levin, M

Davidson, E. and Levin, M. Gene regulatory networks. Proceedings of the National Academy of Sciences, 102 0 (14): 0 4935--4935, 2005

2005
[12]

Safe nonlinear control using robust neural lyapunov-barrier functions

Dawson, C., Qin, Z., Gao, S., and Fan, C. Safe nonlinear control using robust neural lyapunov-barrier functions. In Conference on Robot Learning, pp.\ 1724--1735. PMLR, 2022

2022
[13]

Safe control with learned certificates: A survey of neural lyapunov, barrier, and contraction methods for robotics and control

Dawson, C., Gao, S., and Fan, C. Safe control with learned certificates: A survey of neural lyapunov, barrier, and contraction methods for robotics and control. IEEE Transactions on Robotics, 2023

2023
[14]

D., Nguyen, J., Thakker, R., Alatur, N., Agha-mohammadi, A.-a., and Theodorou, E

Fan, D. D., Nguyen, J., Thakker, R., Alatur, N., Agha-mohammadi, A.-a., and Theodorou, E. A. Bayesian learning-based adaptive control for safety critical systems. In 2020 IEEE international conference on robotics and automation (ICRA), pp.\ 4093--4099. IEEE, 2020

2020
[15]

F., Powell, J

Franklin, G. F., Powell, J. D., Emami-Naeini, A., and Powell, J. D. Feedback control of dynamic systems, volume 4. Prentice hall Upper Saddle River, 2002

2002
[16]

Learning event-triggered control from data through joint optimization

Funk, N., Baumann, D., Berenz, V., and Trimpe, S. Learning event-triggered control from data through joint optimization. IFAC Journal of Systems and Control, 16: 0 100144, 2021

2021
[17]

and Hafstein, S

Giesl, P. and Hafstein, S. Review on computational methods for lyapunov functions. Discrete and Continuous Dynamical Systems-B, 20 0 (8): 0 2291--2331, 2015

2015
[18]

Analysis of event-driven controllers for linear systems

Heemels, W., Sandee, J., and Van Den Bosch, P. Analysis of event-driven controllers for linear systems. International journal of control, 81 0 (4): 0 571--590, 2008

2008
[19]

P., Johansson, K

Heemels, W. P., Johansson, K. H., and Tabuada, P. An introduction to event-triggered and self-triggered control. In 2012 ieee 51st ieee conference on decision and control (cdc), pp.\ 3270--3285. IEEE, 2012

2012
[20]

Sporadic event-based control of first-order linear stochastic systems

Henningsson, T., Johannesson, E., and Cervin, A. Sporadic event-based control of first-order linear stochastic systems. Automatica, 44 0 (11): 0 2890--2895, 2008

2008
[21]

Neural certificates for safe control policies

Jin, W., Wang, Z., Yang, Z., and Mou, S. Neural certificates for safe control policies. arXiv preprint arXiv:2006.08465, 2020

2006 arXiv
[22]

P., Littman, M

Kaelbling, L. P., Littman, M. L., and Moore, A. W. Reinforcement learning: A survey. Journal of Artificial Intelligence Research, 4: 0 237--285, 1996

1996
[23]

Khalil, H. K. Nonlinear systems third edition. Patience Hall, 115, 2002

2002
[24]

Stiff neural ordinary differential equations

Kim, S., Ji, W., Deng, S., Ma, Y., and Rackauckas, C. Stiff neural ordinary differential equations. Chaos: An Interdisciplinary Journal of Nonlinear Science, 31 0 (9), 2021

2021
[25]

Krantz, S. G. and Parks, H. R. The implicit function theorem: history, theory, and applications. Springer Science & Business Media, 2002

2002
[26]

Event-triggered feedback in control, estimation, and optimization

Lemmon, M. Event-triggered feedback in control, estimation, and optimization. Networked control systems, pp.\ 293--358, 2010

2010
[27]

and Liu, Y

Li, F. and Liu, Y. Event-triggered stabilization for continuous-time stochastic systems. IEEE Transactions on Automatic Control, 65 0 (10): 0 4031--4046, 2019

2019
[28]

Lorenz, E. N. Deterministic nonperiodic flow. Journal of atmospheric sciences, 20 0 (2): 0 130--141, 1963

1963
[29]

Stochastic differential equations and applications

Mao, X. Stochastic differential equations and applications. Elsevier, 2007

2007
[30]

and Grebogi, C

Meng, Y. and Grebogi, C. Control of tipping points in stochastic mutualistic complex networks. Chaos: An Interdisciplinary Journal of Nonlinear Science, 31 0 (2), 2021

2021
[31]

Event-based control and signal processing

Miskowicz, M. Event-based control and signal processing. CRC press, 2018

2018
[32]

Ott, E., Grebogi, C., and Yorke, J. A. Controlling chaos. Physical review letters, 64 0 (11): 0 1196, 1990

1990
[33]

Parrilo, P. A. Structured Semidefinite Programs and Semialgebraic Geometry Methods in Robustness and Optimization. California Institute of Technology, 2000

2000
[34]

Automated and formal synthesis of neural barrier certificates for dynamical models

Peruffo, A., Ahmed, D., and Abate, A. Automated and formal synthesis of neural barrier certificates for dynamical models. In International conference on tools and algorithms for the construction and analysis of systems, pp.\ 370--388. Springer, 2021

2021
[35]

Pontryagin, L. S. Mathematical theory of optimal processes. Routledge, 2018

2018
[36]

Learning safe multi-agent control with decentralized neural barrier certificates

Qin, Z., Zhang, K., Chen, Y., Chen, J., and Fan, C. Learning safe multi-agent control with decentralized neural barrier certificates. In International Conference on Learning Representations, 2020

2020
[37]

Reviving a failed network through microscopic interventions

Sanhedrai, H., Gao, J., Bashan, A., Schwartz, M., Havlin, S., and Barzel, B. Reviving a failed network through microscopic interventions. Nature Physics, 18 0 (3): 0 338--349, 2022

2022
[38]

M., Sindhwani, V., Slotine, J.-J

Singh, S., Richards, S. M., Sindhwani, V., Slotine, J.-J. E., and Pavone, M. Learning stabilizable nonlinear dynamics with contraction-based regularization. The International Journal of Robotics Research, 40 0 (10-11): 0 1123--1150, 2021

2021
[39]

Learning certified control using contraction metric

Sun, D., Jha, S., and Fan, C. Learning certified control using contraction metric. In Conference on Robot Learning, pp.\ 1519--1539. PMLR, 2021

2021
[40]

Event-triggered real-time scheduling of stabilizing control tasks

Tabuada, P. Event-triggered real-time scheduling of stabilizing control tasks. IEEE Transactions on Automatic control, 52 0 (9): 0 1680--1685, 2007

2007
[41]

Learning for safety-critical control with control barrier functions

Taylor, A., Singletary, A., Yue, Y., and Ames, A. Learning for safety-critical control with control barrier functions. In Learning for Dynamics and Control, pp.\ 708--717. PMLR, 2020

2020
[42]

Taylor, A. J. and Ames, A. D. Adaptive safety with control barrier functions. In 2020 American Control Conference (ACC), pp.\ 1399--1405. IEEE, 2020

2020
[43]

J., Dorobantu, V

Taylor, A. J., Dorobantu, V. D., Le, H. M., Yue, Y., and Ames, A. D. Episodic learning with control lyapunov functions for uncertain robotic systems. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp.\ 6878--6884. IEEE, 2019

2019
[44]

Tsukamoto, H., Chung, S.-J., and Slotine, J.-J. E. Contraction theory for nonlinear stability analysis and learning-based control: A tutorial overview. Annual Reviews in Control, 52: 0 135--169, 2021

2021
[45]

A geometrical approach to control and controllability of nonlinear dynamical networks

Wang, L.-Z., Su, R.-Q., Huang, Z.-G., Wang, X., Wang, W.-X., Grebogi, C., and Lai, Y.-C. A geometrical approach to control and controllability of nonlinear dynamical networks. Nature communications, 7 0 (1): 0 11323, 2016

2016
[46]

and Louppe, G

Wehenkel, A. and Louppe, G. Unconstrained monotonic neural networks. Advances in neural information processing systems, 32, 2019

2019
[47]

Cybernetics or Control and Communication in the Animal and the Machine

Wiener, N. Cybernetics or Control and Communication in the Animal and the Machine. MIT press, 2019

2019
[48]

Neural network-based event-triggered integral reinforcement learning for constrained h tracking control with experience replay

Xue, S., Luo, B., Liu, D., and Gao, Y. Neural network-based event-triggered integral reinforcement learning for constrained h tracking control with experience replay. Neurocomputing, 513: 0 25--35, 2022

2022
[49]

L yapunov-stable neural control for state and output feedback: A novel formulation

Yang, L., Dai, H., Shi, Z., Hsieh, C.-J., Tedrake, R., and Zhang, H. L yapunov-stable neural control for state and output feedback: A novel formulation. In Salakhutdinov, R., Kolter, Z., Heller, K., Weller, A., Oliver, N., Scarlett, J., and Berkenkamp, F. (eds.), Proceedings o...

2024
[50]

Advancements in mathematical approaches for deciphering deep brain stimulation: A systematic review

Yang, L., Zhang, J., Zhou, S., and Lin, W. Advancements in mathematical approaches for deciphering deep brain stimulation: A systematic review. CSIAM Trans. Life Sci, 1 0 (1): 0 93--133, 2025

2025
[51]

and Miyato, T

Yoshida, Y. and Miyato, T. Spectral norm regularization for improving the generalizability of deep learning. arXiv preprint arXiv:1705.10941, 2017

2017 arXiv
[52]

Neural stochastic control

Zhang, J., Zhu, Q., and Lin, W. Neural stochastic control. In Oh, A. H., Agarwal, A., Belgrave, D., and Cho, K. (eds.), Advances in Neural Information Processing Systems, 2022 a . URL https://openreview.net/forum?id=5wI7gNopMHW

2022
[53]

Sync: Safety-aware neural control for stabilizing stochastic delay-differential equations

Zhang, J., Zhu, Q., Yang, W., and Lin, W. Sync: Safety-aware neural control for stabilizing stochastic delay-differential equations. In The Eleventh International Conference on Learning Representations, 2022 b

2022
[54]

Machine-learning-coined noise induces energy-saving synchrony

Zhang, J., Yang, L., Zhu, Q., Grebogi, C., and Lin, W. Machine-learning-coined noise induces energy-saving synchrony. Physical Review E, 110 0 (1): 0 L012203, 2024 a

2024
[55]

Fessnc: Fast exponentially stable and safe neural controller

Zhang, J., Yang, L., Zhu, Q., and Lin, W. Fessnc: Fast exponentially stable and safe neural controller. In International Conference on Machine Learning, pp.\ 60076--60098. PMLR, 2024 b

2024
[56]

Dynamics of simplicial seirs epidemic model: global asymptotic stability and neural lyapunov functions

Zou, Y., Peng, X., Yang, W., Zhang, J., and Lin, W. Dynamics of simplicial seirs epidemic model: global asymptotic stability and neural lyapunov functions. Journal of Mathematical Biology, 89 0 (1): 0 12, 2024

2024
[57]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

A simple event-based pid controller

arz \'e n, K.-E. A simple event-based pid controller. IFAC Proceedings Volumes, 32 0 (2): 0 8687--8692, 1999

work page 1999

[2] [2]

Stabilization of nonlinear systems using event-triggered output feedback controllers

Abdelrahim, M., Postoyan, R., Daafouz, J., and Ne s i \'c , D. Stabilization of nonlinear systems using event-triggered output feedback controllers. IEEE transactions on automatic control, 61 0 (9): 0 2682--2687, 2015

work page 2015

[3] [3]

D., Xu, X., Grizzle, J

Ames, A. D., Xu, X., Grizzle, J. W., and Tabuada, P. Control barrier function based quadratic programs for safety critical systems. IEEE Transactions on Automatic Control, 62 0 (8): 0 3861--3876, 2016

work page 2016

[4] [4]

Amos, B., Xu, L., and Kolter, J. Z. Input convex neural networks. In International Conference on Machine Learning, pp.\ 146--155. PMLR, 2017

work page 2017

[5] [5]

str \"o m, K. J. and Bernhardsson, B. Comparison of periodic and event based sampling for first-order stochastic systems. IFAC Proceedings Volumes, 32 0 (2): 0 5006--5011, 1999

work page 1999

[6] [6]

Deep reinforcement learning for event-triggered control

Baumann, D., Zhu, J.-J., Martius, G., and Trimpe, S. Deep reinforcement learning for event-triggered control. In 2018 IEEE Conference on Decision and Control (CDC), pp.\ 943--950. IEEE, 2018

work page 2018

[7] [7]

The control of chaos: theory and applications

Boccaletti, S., Grebogi, C., Lai, Y.-C., Mancini, H., and Maza, D. The control of chaos: theory and applications. Physics reports, 329 0 (3): 0 103--197, 2000

work page 2000

[8] [8]

Neural lyapunov control

Chang, Y.-C., Roohi, N., and Gao, S. Neural lyapunov control. In Proceedings of the 33rd International Conference on Neural Information Processing Systems, pp.\ 3245--3254, 2019

work page 2019