Pith. sign in

REVIEW 3 major objections 5 minor 31 references

Learning-enabled Parameter Synthesis for Nonlinear Systems from Signal Temporal Logic

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Neural-network-based parameter learning plus reachability verification gives provable STL guarantees for nonlinear systems.

desk verdict A credible 'learn to propose, verify to certify' pipeline with an independent reachability guarantee; needs a few clarifications about which dynamics the learning step actually uses. read the letter →

arxiv 2607.08899 v2 pith:LQUNUV6L submitted 2026-07-09 eess.SY cs.LGcs.SY

Alex Beaudin , Hanna Krasowski , Eric Palanques-Tost , Calin Belta , Murat Arcak This is my paper · ORCID
classification eess.SYcs.LGcs.SY
keywords parametersynthesissignaltemporallogicneuralODEresidualnetworkreachabilityanalysisset-basedSTLformalverificationnonlinearsystems
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 tries to settle a long-standing trade-off in parameter synthesis from Signal Temporal Logic (STL) specifications: simulation-based methods scale well but offer no formal guarantees, while symbolic methods are sound but struggle with high-dimensional nonlinear systems. The authors propose to get both by using gradient-based training to propose candidate parameters and then a set-based reachability analysis to verify, after the fact, that a chosen parameter vector satisfies the STL formula for every initial condition in a bounded set. The training phase augments the system dynamics with a regularized residual neural network to smooth the non-smooth STL robustness objective, and uses a CVaR-style loss plus a counter-example buffer to focus on worst-case initial conditions. Once training selects a parameter vector, verification computes the reachable set of the true (unaugmented) system and checks the STL specification with set-based semantics; if the check passes, the satisfaction guarantee is as strong as that of symbolic synthesis. The paper demonstrates the method on three systems with up to 18 parameters and reports verified success on all three.

What carries the argument

The load-bearing component is the pairing of a residual neural-network term NN(x;ζ) added to the ODE during training—regularized toward zero so that optimized parameters transfer to the original dynamics—with a posteriori set-based reachability analysis and set-based STL model checking. The surrogate dynamics smooth the rugged STL robustness landscape, while the SlackReLU loss (a conditional value-at-risk objective) and a counter-example buffer direct training toward the worst-case initial conditions. The verification step, using reachable sets and set-based STL semantics, is what converts finite-sample empirical satisfaction into a continuous-domain, all-initial-conditions guarantee.

What would settle it

Take a parameter vector θ* learned by the method (e.g., for the gene regulatory network), set the residual network to zero, and simulate the true system f(x;θ*) from 100,000 random initial conditions drawn uniformly from X0. Compute the discrete-time STL robustness of each trace using the same time step as the original verification; if any of these robustness values is negative, or if the minimum over samples is lower than the verified margin, then the surrogate-to-original transfer has failed and the empirical part of the claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that the scalability of learning-based parameter optimization can be combined with the soundness of model-based verification: a gradient-based learner operating on an augmented, smoothed dynamics produces parameter candidates, and a posteriori reachability analysis lifts these from empirical satisfaction on finitely many samples to continuous-time satisfaction for all initial conditions in X0. The learner minimizes a SlackReLU (conditional value-at-risk) loss over trajectories of the augmented system f(x;θ)+NN(x;ζ), where the residual network is L2-regularized toward zero and trained alongside θ; a counter-example buffer and multi-start initialization mitigate th

Load-bearing premise

The load-bearing premise is that optimizing the augmented dynamics f(x;θ) + NN(x;ζ) with an L2 penalty on the residual produces parameter vectors for which the original system f(x;θ) also satisfies the STL specification—the paper gives no quantitative bound linking the two, only the empirical observation that the residual norm decreases during training.

Editorial extensions

If this is right

  • If the method holds, high-dimensional parameter synthesis from STL specifications—for example, control gains or kinetic rate constants—can be certified for all initial conditions in a bounded set, which is essential for safety-critical applications.
  • The approach closes the gap between fast but unsound simulation-based optimization and sound but poorly scaling symbolic synthesis, offering a practical route to formal guarantees in systems with many parameters.
  • The verification step is independent of the learning step, so any improvement in reachability algorithms or set-based STL checking directly strengthens the guarantees.
  • The joint synthesis of system parameters and control gains (demonstrated on the quadrotor) suggests the method can support co-design of plant and controller from temporal-logic requirements.
  • The training regime—SlackReLU loss, residual smoothing, and counter-example buffer—provides a recipe that could be adapted to other non-smooth objectives beyond STL robustness.

Reading between the lines

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

  • A sharper design might feed the verified robustness—computed from the reachable sets—back into the training loss, closing the loop between learning and verification instead of treating verification as a one-shot filter; this could raise the probability that the learner proposes certifiable parameters.
  • The paper leaves open whether the residual network truly vanishes in general; a formal bound relating the augmented dynamics to the original dynamics (e.g., a Lipschitz-based error estimate) would strengthen the transfer argument, and its absence is the main gap between the empirical and the formal parts of the claim.
  • The method's practical power depends on the tightness of the reachable sets: overly conservative reachability would reject valid parameters, so the choice of reachability tool is as influential as the learner itself.
  • A natural testable extension is to replace the residual network with a regularization on a smoothed STL robustness function (e.g., log-sum-exp) and compare hit rates; the paper's preliminary observation that naive smoothing hurt suggests that the residual-network approach is uniquely suited to this problem, but the reason is not yet fully explained.
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.

Referee Report

3 major / 5 minor

Summary. The paper proposes a two-phase parameter synthesis method for nonlinear ODEs with Signal Temporal Logic (STL) specifications. In the learning phase, a set of parameter candidates is trained by stochastic gradient descent on a CVaR-style robustness objective, using a residual neural ODE appended to the dynamics to smooth the non-convex STL landscape. A candidate is then selected by worst-case robustness on a test set. In the verification phase, the selected parameters are submitted to set-based reachability analysis and set-based STL model checking, which—if successful—provide a continuous-time satisfaction guarantee over all initial conditions. The method is demonstrated on three systems: a PD-controlled quadrotor (12 states, 6 parameters), a gene regulatory network (6 states, 18 parameters), and the Laub-Loomis enzymatic model (7 states, 14 parameters).

Significance. If the method is sound, it offers a scalable alternative to purely symbolic parameter synthesis for nonlinear systems against STL specifications, combining gradient-based learning with formal verification. A clear strength is that the continuous-time guarantee is produced by external, over-approximative reachability tools (CORA and set-based STL), not by the learned surrogate, so the verification step is independent of the training procedure. The paper is also honest about the fact that many random seeds fail to converge, and it makes the code available. The main unresolved issue is the relationship between the augmented dynamics used during training and the original dynamics that are verified; because the residual network does not vanish during training, the paper currently does not establish that the learning phase is optimizing the original problem. This must be clarified with either a transfer argument or explicit evaluation on the original dynamics before the learning component can be considered substantiated.

major comments (3)
  1. [Sec. IV, Eqs. (6)–(9) and Alg. 1] The manuscript never states whether the training objective, the counter-example buffer update, the selection rule (9), and the min-robustness curves in Fig. 2 are evaluated with the original dynamics (1) or the augmented dynamics (7). Equation (6) explicitly imposes (7) for training, and Alg. 1 updates ζ, but the buffer uses the notation x(x_k0, θ) from (1). Since Fig. 4 shows that the residual norm does not vanish (final values around 0.1–0.7), the original and augmented trajectories need not be close over [0,T], and no Lipschitz or error bound is provided. If (9) and the buffer use the augmented dynamics, then θ* is selected for a different system than the one verified, and the reported empirical satisfaction in Fig. 2 does not transfer to the original problem. This is load-bearing for the claim that the learning phase produces candidates for the original problem. Please clarify which
  2. [Sec. IV, Proposition 1] Proposition 1 is incorrect as stated. For the objective in Eq. (8), the minimizer over ν is the α-quantile of the robustness values, not the average of the worst α fraction. At the optimum, L* = -α · (average robustness of the worst α fraction). Therefore a negative L* implies that this average equals -L*/α > 0, not that the worst α fraction has average robustness ν* > 0. The claim that the remaining 1−α fraction have robustness at least ν* is a property of the quantile, not a consequence of L* < 0. This proposition should be corrected; while it does not affect the verification step, it misstates the meaning of the trained objective.
  3. [Sec. VI-A, Fig. 2 and text] The text states that increasing the number of epochs by 200 mitigates non-convergence, e.g., '1 out of 12 seeds converge for the gene network and all converge for the Laub-Loomis model.' This is too thin to support the claim of 'robustly find satisfying parameters.' A single seed out of 12 is a low success rate, and the paper does not report the number of independent training runs or the variance across seeds. Since the method relies on multiple restarts, please report success rates over repeated runs and, if possible, the distribution of test-set robustness over the q candidates.
minor comments (5)
  1. [Sec. IV, Eq. (8)] The argument list in Eq. (8) is written as L(θ,ν,α,{xk0}^{N+M}_{k=1},φ), while Eq. (6) uses L(θ,ν,α,{xk0}^{N+M}_{k=1},φ) and the algorithm uses L(θi,νi,α,X,φ). Please unify the notation.
  2. [Alg. 1] The loop variable e is used in 'e←e+1' but is never initialized or otherwise used. Please remove it or state its purpose.
  3. [Sec. IV, after Eq. (9)] The sentence 'Note that (4) holds when ρ(θ*;Xtest)>0' is ambiguous: (4) is a property of the original system, while (9) may be evaluated on augmented or original dynamics. Please specify which dynamics are used for Xtest and state whether the reported min-robustness values in Fig. 2 are for (1) or (7).
  4. [Fig. 5 caption] The caption says 'Violation regions are in red' for the quadrotor, but no red regions are visible in the printed figure. Please clarify or adjust the figure.
  5. [Sec. VII] The conclusion states that the approach 'consistently identifies valid parameters,' which is stronger than the evidence in Fig. 2, where only a minority of seeds reach satisfaction. Please temper the wording or add quantitative success rates.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: formal guarantee rests on external reachability/STL verification, not on fitted inputs or self-citations.

full rationale

The paper's central guarantee is delegated to an external verification chain: after SGD selects θ* via (9), Sec. IV computes R_[0,T](θ*) with CORA [23] and checks φ with set-based STL [19]; a positive check certifies (5). This step is independent of the authors' fitted values and not derived from a self-citation: [23] and [19] are outside the author set. The learning phase is presented as an empirical candidate generator, not as the producer of the guarantee. Proposition 1 is a tautological property of the SlackReLU loss (8) — negative L* implies positive worst-α average robustness on the drawn samples — and is not used to replace the reachability check. The residual-NODE augmentation (7) and the unstated choice of original vs. augmented dynamics in (9)/buffer are real validation gaps (the residual norm in Fig. 4 does not vanish), but they concern soundness of the learning-to-verification bridge, not equivalence-by-construction of a prediction with its input. Self-citations [9],[10] supply benchmark models and prior gradient-based STL work, not the theorem on which the guarantee rests. Therefore no circular step qualifies under the required standard.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

The guarantee chain rests on two external trusted components (CORA over-approximative reachability and set-based STL checking) and one paper-specific, unanalyzed component (the residual-NODE surrogate). The method's hyperparameters are hand-chosen with minimal sensitivity analysis; per-system learning rates and +200-epoch adjustments were tuned after observing non-convergence. No new physical entities are postulated.

free parameters (6)
  • SlackReLU coefficient α = 0.05
    Small α emulates worst-case robustness; chosen by hand as a stability/fidelity trade-off (Sec. VI). No sensitivity analysis.
  • NODE regularizer λ = 0.1
    Regularizes the residual network toward zero; hand-chosen (Sec. VI). No sensitivity analysis.
  • Per-system learning rate η and epochs E = η=0.05/E=300 (quadrotor), 0.005/500 (gene), 0.002/700 (Laub-Loomis); gene/Laub-Loomis +200 epochs added after observing
    Tuned per system; the paper reports convergence improved after adding 200 epochs, i.e., hyperparameters were adjusted post hoc to obtain the reported results.
  • Seed count q and initialization sample count P = q=16,12,8; P=262144,46656,279936 per system
    Hand-chosen; P up to 279,936 robustness evaluations during initialization. No analysis of how these affect success.
  • Training time step = 1 s
    Single value used throughout; Sec. VII notes coarse steps ease safety but complicate liveness, and finer steps add compute — an acknowledged, unanalyzed trade-off.
  • Batch sizes N, M, buffer size, MLP width, candidate count R = N=32, M=32, buffer=1024, MLP 2×64, R=128
    Fixed hyperparameters with no ablation except the SlackReLU/ReLU/LeakyReLU comparison for the quadrotor.
assumptions (5)
  • domain assumption Set-based reachability analysis [23] and set-based STL model checking [19] (CORA) are sound for the three nonlinear systems
    The continuous-time guarantee (5) rests entirely on this external verification chain; Sec. IV invokes [19], [23], [29] without stating their underlying assumptions.
  • ad hoc to paper The residual-NODE augmented dynamics ẋ=f(x;θ)+NN(x;ζ) is a valid surrogate for optimizing the original system's STL robustness
    No theorem links the augmented training objective to the original dynamics; only empirical decay of ‖ζ‖ (Fig. 4) supports the smoothing claim. This is the load-bearing bridge between learning and verification.
  • domain assumption Discrete-time STL semantics with 1 s sampling during training approximate continuous-time robustness well enough for the optimizer to produce candidates that later pass continuous verification
    Training uses time-discretized traces (Sec. III.B); Sec. VII concedes the step-size trade-off is not systematically analyzed.
  • standard math STL robustness ρ is subdifferentiable and SGD with AdaBelief on the SlackReLU surrogate converges to useful minima
    Standard non-smooth optimization assumption; Fig. 2 shows the assumption frequently fails in practice (gene network: 1/12 seeds converge).
  • standard math ODE (1) has unique solutions on [0,T] for all x₀∈X₀, θ∈Θ
    Existence/uniqueness is needed for the reachable set (2) and trace semantics; not stated explicitly in the paper.
invented entities (1)
  • Residual neural ODE NN(x;ζ) appended to f during training
    purpose: Smooth the non-smooth STL robustness landscape so gradient descent can find satisfying parameters; absent from the final verified model.
    The only supportive evidence is in-paper (Fig. 4: ‖ζ‖ rises then decays). There is no proof that the surrogate preserves the original problem's optima, and no out-of-paper handle. It is a method device, not a physical entity, but it bears the graviton-like burden of doing unverified work inside the claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning-enabled Parameter Synthesis for Nonlinear Systems from Signal Temporal Logic." pith.science (2026). https://pith.science/paper/LQUNUV6L

@misc{pith2026260708899,
  author       = {Pith},
  title        = {Pith review of: Learning-enabled Parameter Synthesis for Nonlinear Systems from Signal Temporal Logic},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LQUNUV6L}},
  note         = {Machine review of arXiv:2607.08899}
}
read the original abstract

Signal Temporal Logic (STL) is increasingly used to describe interpretable objectives and constraints for optimal control and learning methods, especially when no target time series data is available. In this work, we propose to synthesize parameters for nonlinear systems that robustly satisfy continuous-time STL specifications for uncertain initial conditions. To this end, we use gradient-based optimization along with set-based reachability verification to efficiently learn in high-dimensional parameter spaces while providing provable satisfaction guarantees for the optimized parameters. We demonstrate the effectiveness and scalability of our method on three systems with up to 18 parameter dimensions.

Figures

Figures reproduced from arXiv: 2607.08899 by the authors.

Figure 1
Figure 1. Left to right: 1) finding parameters that satisfy the specification [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Training curves with best parameter highlighted for the objective function (top), and minimum robustness of each batch (bottom). (a) Includes a [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Loss landscape for the gene regulatory network. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Norm of residual MLP over training for multiple parameter initial [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Trajectories for various initial conditions. More saturated trajectories indicate higher robustness. The outer black bands indicate the upper and lower [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references

  1. [1]

    Monitoring temporal properties of contin- uous signals,

    O. Maler and D. Nickovic, “Monitoring temporal properties of contin- uous signals,” inFormal Techniques, Modelling and Analysis of Timed and Fault-Tolerant Systems, vol. 3253, 2004, pp. 152–166

    2004

  2. [2]

    Robustness of temporal logic spec- ifications for continuous-time signals,

    G. E. Fainekos and G. J. Pappas, “Robustness of temporal logic spec- ifications for continuous-time signals,”Theoretical Computer Science, vol. 410, no. 42, pp. 4262–4291, 2009

    2009

  3. [3]

    Signal temporal logic neural predictive control,

    Y . Meng and C. Fan, “Signal temporal logic neural predictive control,” IEEE Robotics and Automation Letters, vol. 8, no. 11, pp. 7719–7726, 2023

    2023

  4. [4]

    Control of mobile robots using barrier functions under temporal logic specifications,

    M. Srinivasan and S. Coogan, “Control of mobile robots using barrier functions under temporal logic specifications,”IEEE Transactions on Robotics, vol. 37, no. 2, pp. 363–374, 2021

    2021

  5. [5]

    SpaTiaL: monitoring and planning of robotic tasks using spatio- temporal logic specifications,

    C. Pek, G. F. Schuppe, F. Esposito, J. Tumova, and D. Kragic, “SpaTiaL: monitoring and planning of robotic tasks using spatio- temporal logic specifications,”Autonomous Robots, vol. 47, no. 8, pp. 1439–1462, 2023

    2023

  6. [6]

    Robust online monitoring of signal temporal logic,

    J. V . Deshmukh, A. Donz ´e, S. Ghosh, X. Jin, G. Juniwal, and S. A. Seshia, “Robust online monitoring of signal temporal logic,”Formal Methods in System Design, vol. 51, no. 1, pp. 5–30, 2017

    2017

  7. [7]

    Reinforcement learning with temporal logic rewards,

    X. Li, C.-I. Vasile, and C. Belta, “Reinforcement learning with temporal logic rewards,” inIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2017, pp. 3834–3839

    2017

  8. [8]

    Temporal-logic-based reward shaping for continuing reinforcement learning tasks,

    Y . Jiang, S. Bharadwaj, B. Wu, R. Shah, U. Topcu, and P. Stone, “Temporal-logic-based reward shaping for continuing reinforcement learning tasks,” inAAAI Conference on Artificial Intelligence, 2021, pp. 7995–8003

    2021

Show all 31 references
  1. [9]

    Learning biomolecular models using signal temporal logic,

    H. Krasowski, E. Palanques-Tost, C. Belta, and M. Arcak, “Learning biomolecular models using signal temporal logic,” inLearning for Dynamics and Control Conference (L4DC), 2025, pp. 1365–1377

    2025

  2. [10]

    STL-based optimization of biomolecular neural networks for re- gression and control,

    E. Palanques-Tost, H. Krasowski, M. Arcak, R. Weiss, and C. Belta, “STL-based optimization of biomolecular neural networks for re- gression and control,” inIEEE Conference on Decision and Control (CDC), 2025, pp. 3276–3281

    2025

  3. [11]

    Neu- ral ordinary differential equations,

    R. T. Q. Chen, Y . Rubanova, J. Bettencourt, and D. Duvenaud, “Neu- ral ordinary differential equations,”Advances in Neural Information Processing Systems, vol. 31, 2018

    2018

  4. [12]

    Parameter synthesis through temporal logic specifications,

    T. Dang, T. Dreossi, and C. Piazza, “Parameter synthesis through temporal logic specifications,” inFormal Methods: 20th International Symposium, 2015, pp. 213–230

    2015

  5. [13]

    SAPO: Reachability computation and parameter synthesis of polynomial dynamical systems,

    T. Dreossi, “SAPO: Reachability computation and parameter synthesis of polynomial dynamical systems,” inInternational Conference on Hybrid Systems: Computation and control, 2017, pp. 29–34

    2017

  6. [14]

    A temporal logic approach to modular design of synthetic biological circuits,

    E. Bartocci, L. Bortolussi, and L. Nenzi, “A temporal logic approach to modular design of synthetic biological circuits,” inComputational Methods in Systems Biology, 2013, pp. 164–177

    2013

  7. [15]

    High- performance symbolic parameter synthesis of biological models: a case study,

    M. Demko, N. Bene ˇs, L. Brim, S. Pastva, and D. ˇSafr´anek, “High- performance symbolic parameter synthesis of biological models: a case study,” inInternational Conference on Computational Methods in Systems Biology, 2016, pp. 82–97

    2016

  8. [16]

    Bayesian statistical parameter synthesis for linear temporal properties of stochastic models,

    L. Bortolussi and S. Silvetti, “Bayesian statistical parameter synthesis for linear temporal properties of stochastic models,” inTools and algorithms for the construction and analysis of systems, 2018, pp. 396–413

    2018

  9. [17]

    Set propagation techniques for reachability analysis,

    M. Althoff, G. Frehse, and A. Girard, “Set propagation techniques for reachability analysis,”Annual Review of Control, Robotics, and Autonomous Systems, vol. 4, no. 1, pp. 369–395, 2021

    2021

  10. [18]

    STL model checking of continuous and hybrid systems,

    H. Roehm, J. Oehlerking, T. Heinz, and M. Althoff, “STL model checking of continuous and hybrid systems,” inInternational Sym- posium on Automated Technology for Verification and Analysis, 2016, pp. 412–427

    2016

  11. [19]

    Using four-valued signal temporal logic for incremental verification of hybrid systems,

    F. Lercher and M. Althoff, “Using four-valued signal temporal logic for incremental verification of hybrid systems,” inComputer Aided Verification, 2024, pp. 259–281

    2024

  12. [20]

    Fully automated verification of linear time-invariant systems against signal temporal logic specifications via reachability analysis,

    N. Kochdumper and S. Bak, “Fully automated verification of linear time-invariant systems against signal temporal logic specifications via reachability analysis,”Nonlinear Analysis: Hybrid Systems, vol. 53, 2024

    2024

  13. [21]

    Model predictive control with signal temporal logic specifications,

    V . Raman, A. Donz ´e, M. Maasoumy, R. M. Murray, A. Sangiovanni- Vincentelli, and S. A. Seshia, “Model predictive control with signal temporal logic specifications,” inIEEE Conference on Decision and Control (CDC), 2014, pp. 81–87

    2014

  14. [22]

    Optimization of conditional value- at-risk,

    R. T. Rockafellar and S. Uryasev, “Optimization of conditional value- at-risk,”The Journal of Risk, vol. 2, no. 3, pp. 21–41, 2000

    2000

  15. [23]

    Adaptive parameter tuning for reachability analysis of nonlinear systems,

    M. Wetzlinger, A. Kulmburg, and M. Althoff, “Adaptive parameter tuning for reachability analysis of nonlinear systems,” inInternational Conference on Hybrid Systems: Computation and Control, 2021

    2021

  16. [24]

    Meyer, A

    P.-J. Meyer, A. Devonport, and M. Arcak,Interval Reachability Analysis: Bounding Trajectories of Uncertain Systems with Boxes for Control and Verification. Springer Cham, 2021

    2021

  17. [25]

    Interval signal temporal logic from natural inclusion functions,

    L. Baird, A. Harapanahalli, and S. Coogan, “Interval signal temporal logic from natural inclusion functions,”IEEE Control Systems Letters, vol. 7, pp. 3555–3560, 2023

    2023

  18. [26]

    ARCH-COMP25 Category Report: Continuous and hybrid systems with nonlinear dynamics,

    L. Geretti, M. Althoff, L. Benet, P. Collins, M. Forets, S. Mitsch, I. Patel, M. Perschl, C. Schillinget al., “ARCH-COMP25 Category Report: Continuous and hybrid systems with nonlinear dynamics,” in Proc. of the Int. Workshop on Applied Verification for Continuous and Hybrid S...

    2025

  19. [27]

    Evaluating the robustness of a biochemical network model,

    R. Ghaemi and D. Del Vecchio, “Evaluating the robustness of a biochemical network model,” inIEEE Conference on Decision and Control (CDC), 2007, pp. 615–620

    2007

  20. [28]

    JAX: composable transformations of Python+NumPy programs,

    J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. VanderPlas, S. Wanderman- Milne, and Q. Zhang, “JAX: composable transformations of Python+NumPy programs,” 2018

    2018

  21. [29]

    An introduction to CORA 2015,

    M. Althoff, “An introduction to CORA 2015,” inProc. of the workshop on applied verification for continuous and hybrid systems, 2015, pp. 120–151

    2015

  22. [30]

    Backpropagation through signal temporal logic specifications: Infusing logical structure into gradient-based methods,

    K. Leung, N. Ar ´echiga, and M. Pavone, “Backpropagation through signal temporal logic specifications: Infusing logical structure into gradient-based methods,”The International Journal of Robotics Re- search, vol. 42, no. 6, pp. 356–370, 2023

    2023

  23. [31]

    Control from signal temporal logic specifications with smooth cumulative quantitative semantics,

    I. Haghighi, N. Mehdipour, E. Bartocci, and C. Belta, “Control from signal temporal logic specifications with smooth cumulative quantitative semantics,” inIEEE Conference on Decision and Control (CDC), 2019, pp. 4361–4366

    2019

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.