acceptodds
Under review as a conference paper at ICLR 2027

Statistical Certification of Neural Barrier Functions for invariance and Reach-Avoidance in Learning-Enabled Systems

Abstract

Certifying safety and reach-avoidance for learning-enabled systems with unknown dynamics is challenging, as exact verification is intractable and finite data provide limited evidence. We propose a joint training-and-certification framework that combines multi-step statistical certification of neural barrier functions with reinforcement learning of control policies. Within this framework, PID-barrier feedback jointly trains the policy and barrier to improve safety and performance. After training, we certify probably approximately correct (PAC) multi-step safe invariant and reach-avoid sets using scenario approaches and the trained barrier function. Because the PAC statement bounds the portion of the certified set that may violate the property, rather than certifying individual states, the guarantees are reusable online as long as the system remains within the certified set and the dynamics and controller remain unchanged. The method attains sample complexity, improving on Monte Carlo bounds. Experiments on an inverted pendulum, Van der Pol oscillator, and 10-D Hopfield system demonstrated improved return-safety trade-offs and valid PAC certificates.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Neural Barrier Functions, probably approximately correct, reinforcement learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Jnk7OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a joint training-and-certification framework for learning-enabled control systems with unknown dynamics. The framework combines neural barrier function learning with reinforcement learning to synthesize control policies, followed by scenario-based statistical certification of multi-step safe invariant and reach-avoid sets. During training, a PID-Barrier mechanism adaptively adjusts the constraint multiplier based on empirical barrier violations to improve policy performance and safety. After training, the learned policy and barrier function are fixed, and a scalar scenario optimization problem is solved to establish set-level PAC guarantees. The authors claim an O(1/α)O(1/\alpha) sample complexity, compared with the O(1/α2)O(1/\alpha^2) dependence of standard Monte Carlo bounds. Experiments on an inverted pendulum, a Van der Pol oscillator, and a 10-dimensional Hopfield system are presented to evaluate control performance, safety, and certification efficiency.

Critical strengths

The paper addresses an important research problem concerning safety and reach-avoid certification for learning-enabled systems with unknown dynamics, which is highly relevant to safe reinforcement learning and formal verification.

It presents a unified learning-and-certification framework that integrates policy learning, neural barrier functions, and statistical certification in a clear and well-organized manner. The separation of training and certification, together with the use of an independent certification dataset, is a reasonable design choice. Furthermore, the framework extends existing one-step PAC controlled invariance methods to multi-step invariance and reach-avoidance properties, broadening their applicability.

The experimental results are also promising, suggesting that PID-Barrier can improve the trade-off between control performance and safety, while potentially producing more certifiable policy-barrier pairs for reach-avoidance tasks.

Critical weaknesses

-The proposed framework mainly combines several established techniques, including neural barrier learning, PPO, PID-Lagrangian optimization, and scenario-based certification. While their integration may be practically useful, the manuscript does not sufficiently clarify the technical novelty of the individual components or explain why their combination constitutes a substantial methodological advance.

-Definition 1 introduces infinite-horizon safe invariance, whereas Definition 3 only considers a finite horizon. The distinction between finite-horizon PAC safety and classical invariant-set guarantees should be emphasized. Another concern is that the conditional guarantee 1−α/μS(S~)1-\alpha/\mu_S(\widetilde S) can become vacuous when the certified set has a small measure. Since the authors acknowledge that computing this measure may be challenging, the manuscript should explain how meaningful, non-vacuous guarantees are established in practice.

-The paper emphasizes an O(1/α)O(1/\alpha) sample complexity for scenario-based certification, compared with O(1/α2)O(1/\alpha^2) for Monte Carlo methods based on Hoeffding's inequality. However, these bounds may correspond to different statistical objectives. The O(1/α2)O(1/\alpha^2) dependence typically arises from estimating an unknown probability with additive accuracy, whereas one-sided violation-probability certification can also achieve O(1/α)O(1/\alpha) dependence under appropriate conditions. Although the data-dependent nature of the certified sets makes the proposed scenario formulation different from certification of a fixed set, the comparison with Hoeffding-based Monte Carlo estimates alone does not fully establish a general sample-efficiency advantage. The authors should clarify whether the compared methods provide equivalent guarantees and consider stronger statistical certification baselines.

-Experimental evidence is insufficient to fully support the main claims. Although the experiments cover three benchmark systems, the evaluation remains relatively limited for a framework claiming improved scalability and certification efficiency. First, the certification results do not consistently demonstrate an advantage over existing barrier-learning methods. For example, in the inverted pendulum invariance task, MF-NBC and Lag-Barrier achieve larger certified volumes than the proposed method. This weakens the claim that PID-Barrier consistently improves certification quality. Second, the main certification comparison is concentrated on the inverted pendulum, while detailed results for the other systems are largely deferred to the appendix. More comprehensive certification results across all benchmarks would strengthen the evaluation.

Minor Comments:

i. The content in Lines 63–66 of the Introduction appears to overlap with the Abstract.

ii. The methodological novelty over Wu et al. (2026) remains unclear, particularly whether the main advancement is simply replacing convex optimization with neural learning and verification. The authors should clarify the technical contributions beyond this extension.

iii. The Notations paragraph in Line 114 would be more appropriately placed at the beginning of the Preliminaries section.

iv. In Remark 1, it is unclear why Monte Carlo techniques are suggested for probability estimation, as the proposed framework does not appear to adopt this approach.

v. The piecewise construction of the neural barrier function in Eq. (4) appears to be an important design choice, but its motivation and role in training are not sufficiently explained. The authors should elaborate on why this construction is necessary and how it benefits learning and certification, rather than focusing primarily on the relatively straightforward loss functions and dataset construction.

vi. In Line 349, “equation 11” should be changed to “equation (11)”.

vii. Figure 1 is too small and poorly scaled, making its contents difficult to read. The authors should adjust its size and proportions to improve readability.

viii. Could the authors provide the link to the anonymous repository mentioned in the Reproducibility Statement?

Overall, I believe this work has merit, and I would be inclined to raise my rating if the authors could more clearly articulate the technical novelty and provide more comprehensive experimental evaluations.

Questions for authors
  1. What is the main theoretical novelty of the multi-step PAC certification compared with Wu et al. (2026)? Does extending one-step constraints to multi-step trajectories require fundamentally new scenario optimization arguments, or can the proposed guarantees be derived directly from existing results?

  2. How does the proposed PID-Barrier update differ technically from combining existing neural barrier learning with PID-Lagrangian optimization? Can the authors provide additional ablations demonstrating the independent contributions of the P, I, and D components?

  3. In Table 1, existing barrier-learning methods achieve larger certified volumes than PID-Barrier for the invariance task. What evidence supports the claim that PID-Barrier improves certification quality rather than primarily improving empirical control performance?

4 comments
AuthorsOpenReview ↗

Response 1

We thank the reviewer for these constructive comments. We will address all the minor comments in the revised manuscript, and add experiments on additional high-dimensional systems in the final version to better demonstrate the scalability of the proposed method.

Weaknesses 1 and 3, and Question 1. Technical contributions and sample efficiency.

We thank the reviewer for these important questions.

(1) Our contribution is a joint training-and-certification framework that establishes PAC multi-step safe invariant and reach-avoid sets for learned controllers in black-box systems. Within this framework, empirical barrier violations guide joint policy--barrier learning, and independent trajectory data are used to statistically certify the resulting learned controller. This connects policy learning with the construction of certified operating regions under unknown dynamics.

Relative to (Wu et al.(2026)), we extend the certification objective from the existence of a suitable input for one-step controlled invariance to the multi-step behavior of the actual learned controller, including safe target reaching. This requires relating trajectory-level barrier violations to the desired specifications. In particular, Theorems 1 and 2 establish how a calibrated slack determines the barrier level sets for which the corresponding finite-horizon property holds, up to a bounded measure of exceptional states. This construction also accommodates positive slack in the admissible range, allowing useful certificates even when the original barrier inequalities are imperfectly satisfied.

A further distinction lies in the underlying learning and certification paradigms. Wu et al. (2026) adopts a uniform-learning approach, whereas our work follows a learning-then-verification paradigm. Specifically, we first learn the policy and neural barrier and then use independent trajectory samples to statistically certify the multi-step safety and reach-avoid properties of the resulting learned controller. This separation allows us to establish PAC guarantees for the actual learned networks without requiring uniform guarantees over the entire function class. As a result, unlike Wu et al. (2026), our framework does not impose a convexity requirement on the learned functions used to characterize invariant and reach-avoid sets. It accommodates general, potentially nonconvex neural barrier functions, offering greater flexibility in representing the geometry of certified sets.

(2) Regarding sample efficiency, we agree that comparing our framework only with a simple Monte Carlo baseline does not sufficiently demonstrate its advantages relative to established statistical certification methods. We will therefore include an exact one-sided binomial certification baseline based on the Clopper–Pearson confidence bound.

For a fixed candidate set selected independently of the validation data, the exact one-sided Clopper-Pearson confidence bound provides a natural baseline. Suppose that kk violations are observed among NN independent validation trajectory. The corresponding upper confidence bound at confidence level 1−β1-\beta is Beta⁡−1(1−β;k+1,N−k).\operatorname{Beta}^{-1}(1-\beta;k+1,N-k). In the zero-failure regime k=0k=0, this reduces to 1−β1/N.1-\beta^{1/N}. Consequently, certifying that the violation probability is at most α\alpha requires N≥ln⁡(1/β)−ln⁡(1−α)=O(1/α).N\geq \frac{\ln(1/\beta)}{-\ln(1-\alpha)}=O(1/\alpha). for a fixed confidence level. Thus, we acknowledge that our O(1/α)O(1/\alpha) sample-complexity dependence does not constitute an asymptotic improvement over exact binomial certification in the zero-failure regime. However, when violations are observed, the required sample size depends on their frequency and the desired accuracy. In the general nonzero-failure regime, where the violation probability must be estimated to an additive accuracy of order α\alpha, the usual concentration-based sample complexity is O(1/α2)O(1/\alpha^2) for fixed confidence. The exact binomial bound can have different scaling in rare-event regimes, depending on the number of observed violations.

However, the comparison above concerns a fixed certification event and does not capture the main advantage of our framework, which lies in its treatment of data-dependent set selection. Specifically, our calibrated slack ξ∗\xi^\ast is obtained by solving a scenario optimization problem using the certification data. The scenario theorem then provides an out-of-sample probabilistic guarantee for this data-dependent solution, accounting for its dependence on the sampled trajectories. Under the assumptions of the scenario theorem, our resulting sample-complexity bound retains an O(1/α)O(1/\alpha) dependence. Furthermore, Theorems 1 and 2 establish a deterministic connection between the calibrated slack and a family of barrier level sets. Consequently, a single certification procedure supports PAC guarantees for all admissible barrier thresholds, without requiring separate trajectory validation for each candidate set.

AuthorsOpenReview ↗

Response 2

In contrast, applying an ordinary Clopper-Pearson bound to a candidate set selected using the same validation data does not automatically preserve its nominal confidence guarantee. Independent validation or an appropriate simultaneous inference procedure is needed to account for data-dependent selection. Our framework therefore provides an integrated approach to selection-aware PAC certification of a family of multi-step operating sets.

Finally, we emphasize that the primary contribution of our work is not a new statistical confidence bound, but a framework for PAC certification of multi-step safe invariant and reach-avoid sets for black-box systems under learned controllers. To the best of our knowledge, this is the first work to establish such guarantees within a learning-then-verification framework using trajectory-based scenario optimization. We will clarify these distinctions and include the Clopper-Pearson baseline in the revised manuscript to better demonstrate the statistical contributions of our framework.

Weakness 2: Finite-horizon versus infinite-horizon guarantees; non-vacuous probability bounds for small certified sets; rigorous computation of certified-set measures.

We thank the reviewer for this insightful question.

Definition 1 introduces classical safe invariant sets, which require trajectories starting in the set to remain there for all future time steps. Computing or certifying such sets is generally challenging and can be intractable, particularly for black-box systems with unknown dynamics. In contrast, Definition 3 establishes a finite-horizon PAC guarantee, bounding the proportion of initial states within the certified set from which trajectories remain in the set over the next kmax⁡k_{\max} steps. Thus, our guarantee concerns finite-horizon invariance and does not, by itself, imply classical infinite-horizon invariance. We will clarify this distinction in the revised manuscript and explicitly emphasize the finite-horizon nature of our certification.

Regarding the conditional probability lower bound, we acknowledge that 1−αμS(S~ζ)1-\frac{\alpha}{\mu_S(\widetilde S_\zeta)} can become non-positive when the certified set has a sufficiently small measure. Although the bound remains mathematically valid, it becomes uninformative in this case. In particular, it is positive whenever μS(S~ζ)>α\mu_S(\widetilde S_\zeta)>\alpha. Our framework provides several mechanisms to improve the usefulness of this guarantee. First, during policy and barrier learning, the volume regularization term LvolL_{\rm vol} encourages the neural barrier to produce larger candidate certified sets. Second, after training, our framework allows the certification threshold ζ\zeta to be varied within its admissible range, yielding a family of certified sets with different sizes without retraining the learned networks. Third, increasing the certification sample size allows us to obtain a smaller violation parameter α\alpha while maintaining the same confidence level, thereby improving the conditional probability lower bound. Furthermore, users can specify a desired minimum conditional satisfaction probability ϵ∈(0,1−α)\epsilon\in(0,1-\alpha) and require the certified sets to satisfy

\geq \epsilon,$$ or, equivalently, $$\mu_S(\widetilde S_\zeta) \geq \frac{\alpha}{1-\epsilon}.$$ This provides an explicit criterion for balancing the size of the certified region against the desired probability guarantee. This requirement can be incorporated into Definitions 3 and 4 by replacing the condition $\mu_S(\widetilde S_\zeta)>0$ with $$1-\frac{\alpha}{\mu_S(\widetilde S_\zeta)} \geq \epsilon.$$ Importantly, this modification does not require changing our scenario optimization procedure or its theoretical analysis, since Theorems 1 and 2 already establish the corresponding conditional probability bounds whenever the certified set has positive measure. It simply restricts attention to certified sets for which these bounds are sufficiently informative. Finally, as discussed in Remark 1, computing the exact measure of a neural-network-defined certified set can be challenging. Rigorous lower bounds on this measure can, in principle, be obtained using more advanced techniques, such as interval-arithmetic-based set analysis. However, developing or comparing such volume-computation methods is beyond the scope of this work. Our focus is on establishing PAC guarantees for multi-step safety and reach-avoid properties, rather than on the exact computation of certified-set measures.
AuthorsOpenReview ↗

Response 3

Question 2: How does PID-Barrier differ from combining barrier learning with PID-Lagrangian optimization, and what do the P, I, and D ablations show?

We thank the reviewer for this insightful question. Our PID-Barrier method builds on the PID-Lagrangian principle but specializes its feedback mechanism to the joint learning of the control policy and neural barrier function. In this setting, the barrier condition provides the constraint signal guiding policy improvement. However, this signal is inherently non-stationary because both the policy and barrier networks evolve during training. Our motivation for introducing PID feedback is to adaptively regulate these evolving barrier violations and improve the quality and stability of the policy–barrier pairs subsequently used for certification.

Specifically, we apply PID feedback to the empirical barrier violation em=ρθLsafe(m)(ϕ),e_m=\rho_\theta L_{\rm safe}^{(m)}(\phi), where ρθ\rho_\theta is the PPO importance ratio. The multiplier is updated according to Im=max⁡0,Im−1+em,I_m = \max\\{0,I_{m-1}+e_m\\}, Dm=max⁡0,em−em−1,D_m = \max\\{0,e_m-e_{m-1}\\}, ν(m+1)=max⁡0,KPem+KIIm+KDDm.\nu^{(m+1)}=\max\\{0,K_Pe_m+K_II_m+K_DD_m\\}. Here, the proportional term responds to the current barrier violation, the integral term accumulates persistent violations, and the derivative term responds to increases in the violation signal. The non-negative projection maintains the one-sided penalty associated with the barrier constraint. Thus, PID-Barrier uses the temporal evolution of the barrier violations to adapt the policy optimization pressure, rather than relying solely on the accumulated constraint violation in a conventional Lagrangian update.

To investigate the respective contributions of the PID components, Appendix I compares the P, PI, and PID-Barrier variants, together with the Lag-P baseline, in terms of return, control cost, and verification performance. The results show that the benefit of incorporating the integral and derivative feedback is particularly evident in the Van der Pol system. Its nonlinear dynamics induce substantial trajectory fluctuations, making policy updates more susceptible to short-term disturbances. PID-Barrier exhibits better return–cost behavior in this setting, with its adaptive multiplier helping the policy recover from short-term performance degradation. We further evaluate the resulting verification errors in both invariant and reach-avoid tasks, where PID-Barrier maintains stable results within the prescribed verification bounds.

These ablation results provide empirical evidence that the PID components can improve the adaptation of barrier-constraint feedback during joint policy–barrier learning, particularly in challenging nonlinear systems. We will clarify the methodological distinction from conventional PID-Lagrangian optimization and explain the ablation results more explicitly in the revised manuscript.

AuthorsOpenReview ↗

Response 4

Weakness 4 and Question 3: What evidence supports improved certification quality when existing barrier-learning methods obtain larger certified volumes?

We thank the reviewer for this insightful observation. We agree that certified-set volume is an important measure of certification performance. However, a larger certified set does not necessarily translate into better overall performance of the learned controller. Our training procedure seeks to balance task reward and barrier-condition satisfaction, while the volume regularization term encourages larger candidate certified sets. Accordingly, we evaluate the learned methods based on several complementary criteria, including task return, control cost, the validity of the resulting certificates, and the estimated measures of the certified sets. Our objective is to obtain controllers that achieve good task performance while admitting useful safety and reach-avoid certificates, rather than maximizing certified-set volume in isolation.

The inverted pendulum results illustrate this trade-off. For invariance, MF-NBC obtains a larger estimated certified-set measure than PID-Barrier (0.551 versus 0.461). However, PID-Barrier achieves substantially higher return (179.41 versus 100.47) and lower control cost (0.042 versus 0.250), while both methods yield valid certificates across all five seeds. Compared with Lag-Barrier, PID-Barrier achieves a slightly larger estimated certified-set measure (0.461 versus 0.458), together with higher return and lower cost. For reach-avoidance, although PID-Barrier produces a smaller estimated set measure than Lag-Barrier (0.405 versus 0.488), it achieves higher return (361.49 versus 348.37) and lower cost (0.040 versus 0.064), with valid certificates across all five seeds. In contrast, MF-NBC obtains a valid reach-avoid certificate in only one of the five seeds.

The Van der Pol results provide further evidence of the benefits of our approach. For invariance, PID-Barrier achieves a larger estimated certified-set measure than Lag-Barrier (0.624 versus 0.588), with valid certificates in all five seeds compared with four out of five for Lag-Barrier. It also achieves substantially higher return (189.79 versus 128.23) and lower control cost (0.024 versus 0.230). For reach-avoidance, PID-Barrier also obtains a substantially larger estimated certified-set measure (0.586 versus 0.365), with valid certificates in all five seeds for both methods. The corresponding certification slack is also more concentrated across random seeds. Although Lag-Barrier achieves slightly higher return and lower cost in this case, the results demonstrate a clear improvement in the size of the certified operating region while maintaining competitive control performance.

Overall, these results show that the advantage of PID-Barrier is not necessarily a larger certified set in every task, but a more favorable balance between control performance and certification outcomes. In particular, the Van der Pol results demonstrate that our learning procedure can improve both the certified-set measure and control performance, while the inverted pendulum results illustrate the trade-offs involved. We will make these distinctions more explicit and include the results across benchmarks in the revised main comparison.

Reviewer_wBPcOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents a joint learning and statistical certification framework for finite-horizon invariance and reach-avoidance in black-box dynamical systems. During training, a PPO-based controller and a neural barrier function are learned jointly, with a PID-style multiplier adapting the influence of barrier violations on policy optimization. After training, the controller and barrier are frozen, and certification is performed using independently sampled initial states and their trajectories. The certification stage formulates a scalar scenario optimization problem and derives PAC guarantees on the proportion of initial states within a certified set that satisfy the corresponding multi-step specification. These guarantees are defined with respect to a uniform-volume reference measure and apply to the certified set as a whole, rather than to individual states. The formulation also supports a family of nested certified sets through threshold adjustment without retraining. The authors further claim a sample-complexity advantage over Monte Carlo certification based on Hoeffding bounds. The framework is evaluated on an inverted pendulum, a Van der Pol oscillator, and a 10-dimensional Hopfield system, with comparisons of learning performance, certification outcomes, and selected design parameters.

Critical strengths
  1. The main strength of the paper is its attempt to combine policy learning with an independent post-training certification procedure. Treating each complete trajectory as a scenario is well suited to the multi-step specifications, since the constraints along a trajectory can be considered jointly without assuming independence between successive states.

  2. Another useful aspect is that, once the policy and neural barrier are fixed, the certification problem is formulated in terms of a scalar optimization variable. This makes the scenario approach applicable without requiring the learned neural barrier itself to be convex in its parameters. The resulting family of nested certified sets, obtained without retraining the policy or barrier, also offers some practical flexibility. I consider this integration relevant for learning-enabled systems with unknown dynamics. Although the individual components build substantially on existing methods, combining them into a training-and-certification workflow is a positive aspect of the work.

Critical weaknesses
  1. My main concern is the gap between the theoretical guarantees and their numerical implementation. Theorems 1 and 2 rely on the true measure of the certified set, μS(S~ζ)\mu_S(\widetilde S_\zeta), whereas Algorithm 2 substitutes an empirical estimate, p^ζ\hat p_\zeta, obtained from the initial-state samples also used for certification. Although Appendix E mentions additional samples for volume estimation, it remains unclear how estimation uncertainty and data dependence are accounted for. Without a valid confidence correction, the reported probability lower bounds do not automatically inherit the guarantees established by the theorems. This is particularly concerning because these bounds are central to the paper's contribution.

  2. The claimed improvement from O(1/α2)O(1/\alpha^2) to O(1/α)O(1/\alpha) is primarily established against a Hoeffding-based Monte Carlo bound, rather than Monte Carlo certification in general. For a fixed event with no observed failures, the exact one-sided Clopper–Pearson bound also achieves O(log⁡(1/β)/α)O(\log(1/\beta)/\alpha) sample complexity. The potential advantage of scenario optimization for data-dependent set selection is a separate contribution that should be distinguished from fixed-set validation and supported by a more appropriate comparison.

  3. Several reported results appear inconsistent with the experimental settings. Appendix E specifies a maximum episode length of 200 and per-step rewards bounded above by 1, yet Appendix G, Table 3 reports reach-avoid returns of 361.49 for the inverted pendulum and 380.97 for Hopfield. These values require clarification regarding the evaluation horizon or return definition. Furthermore, for Hopfield invariance, the validity condition requires ζ<0.1\zeta<0.1, while Table 5 reports a median ζ=0.1165\zeta=0.1165 with only 2/5 valid runs. It is unclear how the reported certified-set measure and theoretical probability bound relate to the valid runs. These ambiguities affect the interpretation of the experimental guarantees.

  4. The paper combines established ideas from barrier-guided policy learning, scenario certification, and PID multiplier updates. The Lag-Barrier comparison and PID ablations provide useful evidence, but they do not fully separate the benefit of barrier-based feedback from that of using PID updates. A matched PID-Lagrangian baseline using the same task-level cost as Lag-Cost, with comparable architectures, training budgets, and tuning effort, would help isolate this distinction. My concern is the evidence for the added benefit of the combination, rather than the use of established components itself.

  5. The PAC guarantee already concerns finite-horizon trajectories, but with initial states measured under the fixed reference measure μS\mu_S. It does not automatically provide the same bound for a particular observed state or for the conditional state distribution encountered during deployment. The set-level statement remains valid when the controller and dynamics are unchanged; however, its reuse does not by itself establish a new conditional safety bound at each time step or an infinite-horizon guarantee. The claimed operational benefit of reuse should be distinguished from the continued validity of the original set-level statement.

Questions for authors

Q1. Theorems 1 and 2 depend on the true measure of the certified set, while Algorithm 2 uses an empirical estimate. Which samples were used to compute the certified-set measures reported in the tables? Were they independent of the scenario data and threshold selection? More importantly, how is volume-estimation uncertainty incorporated into the final 1−β1-\beta confidence guarantee? Could the authors provide a valid confidence-adjusted derivation of the reported probability lower bounds?

Q2. How does the proposed O(1/α)O(1/\alpha) sample-complexity advantage compare with exact one-sided binomial confidence bounds, such as Clopper–Pearson, in the zero-failure regime for a fixed certification event? What additional advantage does the scenario formulation provide when accounting explicitly for data-dependent set selection? A comparison that separates these two settings would help establish the claimed statistical contribution.

Q3. How are the reach-avoid returns exceeding 200 in Appendix G, Table 3 obtained under the reward definition and episode-length limit specified in Appendix E? Also, for Hopfield invariance in Table 5, which runs are used to report the certified-set measure and theoretical probability lower bound, given that the median certification slack exceeds the required validity threshold and only 2/5 runs are valid? These clarifications are necessary to assess the reported experimental guarantees.

Q4. What evidence separates the benefit of barrier-based feedback from the benefit of PID multiplier updates? In particular, how does PID-Barrier compare with PID-Lagrangian using the same task-level cost as Lag-Cost under matched architectures, training budgets, and comparable tuning effort?

4 comments
AuthorsOpenReview ↗

Response 1

Weakness 1 and Q1: Which samples were used for the certified-set measures, and how are volume-estimation uncertainty and data dependence accounted for?

We thank the reviewer for raising this important question. As discussed in Remark 1, computing the exact reference measure of a certified set can be challenging because the set is represented by a neural barrier function. We therefore use Monte Carlo sampling to estimate its measure in our experiments.

We clarify that two independent datasets are used for different purposes. The first dataset consists of initial states and their corresponding trajectories, which are used in scenario optimization to obtain the calibrated slack and construct the certified set. The second dataset, independently sampled from the reference measure, is used exclusively to estimate the measure of the resulting certified set. Thus, the volume-estimation samples are independent of both the scenario data and the threshold selection.

Importantly, Theorems 1 and 2 establish the PAC guarantees in terms of the true reference measure of the certified set. Specifically, with confidence at least 1−β1-\beta over the scenario dataset, the reference measure of the initial states in the certified set that violate the corresponding multi-step property is bounded by α\alpha. Consequently, the conditional probability of satisfying the property, given that the initial state belongs to S~ζ\widetilde S_\zeta, is bounded below by 1−αμS(S~ζ),1-\frac{\alpha}{\mu_S(\widetilde S_\zeta)}, where μS(S~ζ)\mu_S(\widetilde S_\zeta) denotes the true measure of the certified set.

In our experiments, we substitute the Monte Carlo estimate of this measure into the theoretical lower bound to obtain its numerical approximation. We emphasize that the sampling uncertainty associated with this volume estimate is not incorporated into the theoretical PAC confidence parameter 1−β1-\beta. Accordingly, the reported values are empirical estimates of the theoretical lower bounds, rather than separately confidence-adjusted lower bounds.

We will clarify this distinction in the revised manuscript, explicitly distinguishing the theoretical PAC guarantees from their numerical evaluation.

Weakness 2 and Q2: Sample complexity and set selection.

We thank the reviewer for raising this important question. We agree that the comparison with exact binomial confidence bounds should distinguish fixed-event certification from data-dependent set selection.

For a fixed candidate set selected independently of the validation data, the exact one-sided Clopper-Pearson confidence bound provides a natural baseline. Suppose that kk violations are observed among NN independent validation trajectory. The corresponding upper confidence bound at confidence level 1−β1-\beta is Beta⁡−1(1−β;k+1,N−k).\operatorname{Beta}^{-1}(1-\beta;k+1,N-k). In the zero-failure regime k=0k=0, this reduces to 1−β1/N.1-\beta^{1/N}. Consequently, certifying that the violation probability is at most α\alpha requires N≥ln⁡(1/β)−ln⁡(1−α)=O(1/α)N\geq \frac{\ln(1/\beta)}{-\ln(1-\alpha)}=O(1/\alpha) for a fixed confidence level. Thus, we acknowledge that our O(1/α)O(1/ \alpha) sample-complexity dependence does not constitute an asymptotic improvement over exact binomial certification in the zero-failure regime. However, when violations are observed, the required sample size depends on their frequency and the desired accuracy. In the general nonzero-failure regime, where the violation probability must be estimated to an additive accuracy of order α\alpha, the usual concentration-based sample complexity is O(1/α2)O(1/ \alpha^2) for fixed confidence. The exact binomial bound can have different scaling in rare-event regimes, depending on the number of observed violations.

However, the comparison above concerns a fixed certification event and does not capture the main advantage of our framework, which lies in its treatment of data-dependent set selection. Specifically, our calibrated slack ξ∗\xi^\ast is obtained by solving a scenario optimization problem using the certification data. The scenario theorem then provides an out-of-sample probabilistic guarantee for this data-dependent solution, accounting for its dependence on the sampled trajectories. Under the assumptions of the scenario theorem, our resulting sample-complexity bound retains an O(1/α)O(1/\alpha) dependence. Furthermore, Theorems 1 and 2 establish a deterministic connection between the calibrated slack and a family of barrier level sets. Consequently, a single certification procedure supports PAC guarantees for all admissible barrier thresholds, without requiring separate trajectory validation for each candidate set.

AuthorsOpenReview ↗

Response 2

In contrast, an ordinary Clopper-Pearson bound applied to a fixed candidate set does not automatically provide simultaneous guarantees for a family of sets selected using the same validation data. Such guarantees would require independent validation data or an appropriate simultaneous inference procedure to account for data-dependent selection. Our scenario formulation therefore provides an integrated approach to selection-aware PAC certification, supporting a family of multi-step operating sets through a single scenario-based calibration.

Finally, we emphasize that the primary contribution of our work is not a new statistical confidence bound, but a framework for PAC certification of multi-step safe invariant and reach-avoid sets for black-box systems under learned controllers. To the best of our knowledge, this is the first work to establish such guarantees within a learning-then-verification framework using trajectory-based scenario optimization. We will clarify these distinctions and include the Clopper-Pearson baseline in the revised manuscript to better demonstrate the statistical contributions of our framework.

Weakness 3 and Q3: Experimental reporting.

We thank the reviewer for carefully identifying these issues. We have reexamined the reward definition and the run-wise certification results and will clarify and correct the corresponding descriptions and tables.

Regarding the reach-avoid returns in Appendix G, Table~3.

In the policy net and value net, we use γ=0.99\gamma=0.99. In our all present value, we all use the episodic return with γ=1\gamma=1. We identified a typographical error in the reward definition in Appendix E (Lines 925--929): the reward for entering the target set was incorrectly specified as 11, whereas the actual reward used in our experiments is

\begin{cases} 1-c_u\lVert u\rVert_2^2, & \text{if } f(x,u)\in\mathcal{S}\setminus\mathcal{T},\\\\ 2-c_u\lVert u\rVert_2^2, & \text{if } f(x,u)\in\mathcal{T},\\\\ -100, & \text{if } f(x,u)\notin\mathcal{S}. \end{cases}$$ The additional reward for entering the target set encourages the controller to reach the target while maintaining safety, rather than remaining indefinitely in a safe but non-target region. Since the per-step reward is bounded above by $2$, the maximum undiscounted cumulative reward over a 200-step episode is $2\times 200=400$. For clarity, all episodic returns reported in our tables are undiscounted (i.e., computed with $\gamma=1$). The discount factor $\gamma=0.99$ is used separately in PPO to compute the discounted returns for policy and value network optimization. Therefore, the returns exceeding $200$ in Appendix G, Table 3 are consistent with both the reward function used in our experiments and the 200-step episode limit. We have corrected the reward definition in Appendix E. ### Regarding the Hopfield invariance results in Table~5. we clarify the convention used to report the results. Each entry in a $\zeta$ or $\zeta'$ column is reported as $\mathrm{median}\,[\min,\max]$ over five random seeds. In contrast, the quantities $P$, $\mathrm{Theo.}$, and $\mathrm{Safe}$ are evaluated using the run corresponding to the median certification slack, as specified in the table caption. For PID-Barrier in the Hopfield invariance task, the five certification slack values are $$0.160830,\quad 0.093405,\quad 0.170512,\quad 0.094454,\quad 0.116524.$$ Accordingly, the reported summary is $$0.1165\,[0.0934,\,0.1705],$$ where $0.1165$ is the median and $0.0934$ and $0.1705$ are the minimum and maximum, respectively. Under the experimental setting $\lambda=0.9$ and $C=-1$, the validity condition in Theorem 1 requires $$\xi^* < (\lambda-1)C = 0.1.$$ Only two of the five runs satisfy this condition, which explains the reported validity count of $2/5$. In particular, the run corresponding to the median slack (seed 4), with $\xi^*=0.116524$, does not satisfy the validity condition. In the original table, the reported $P$ and $\mathrm{Theo.}$ values were calculated using this median-slack run. Its estimated certified-set measure is $P=0.0309$. Substituting this value and $\alpha=0.0159829291$ into the expression $$\mathrm{Theo.} = 1-\frac{\alpha}{P} = 1-\frac{0.0159829291}{0.0309}\approx0.482753$$ gives the value $0.483$ reported in the table. Although this numerical calculation is correct, the median-slack run does not satisfy the validity condition of Theorem 1. Consequently, this value cannot be interpreted as a theorem-supported probability lower bound. We acknowledge this inconsistency and will replace the $P$ and $\mathrm{Theo.}$ entries with ``--'' throughout the invariant-task block of Table 5, since the median-slack run for each method in this block fails the corresponding validity condition. The five-seed slack summaries and validity counts will be retained. The $\mathrm{Safe}$ entries, which report empirical probabilities rather than theorem-supported bounds, will also be retained.
AuthorsOpenReview ↗

Response 3

We have also reexamined Tables 1 and 4. The median-slack runs in Table 1 satisfy the corresponding theorem conditions; therefore, the reported certification results in that table remain unchanged. In Table 4, however, the median-slack runs of PPO and MF-NBC in the invariant task do not satisfy the required validity condition. We will therefore replace their PP and Theo.\mathrm{Theo.} entries with ``--'' as well.

These corrections ensure that the reported theoretical probability lower bounds are associated only with runs satisfying the corresponding theorem conditions. We will also clarify the distinction between the five-seed summary statistics and the individual runs used to report the certification quantities.

Weakness 4 and Q4: What evidence separates barrier feedback from PID multiplier updates, including a matched PID-Lagrangian baseline?

We thank the reviewer for this insightful question. Our motivation for PID-Barrier is to make policy improvement respond directly to the barrier conditions used in subsequent certification. Specifically, the learned barrier provides transition-level information about constraint violations, while PID feedback adaptively regulates the strength of this signal as the policy and barrier networks are jointly updated. This combination aims to produce policy–barrier pairs with a favorable return–safety trade-off and useful certification results.

Our method builds on the PID-Lagrangian principle but applies the PID update to the empirical barrier violation em=ρθLsafe(m)(ϕ),e_m=\rho_\theta L_{\rm safe}^{(m)}(\phi), rather than a task-level safety cost. The proportional term responds to the current barrier violation, the integral term accumulates persistent violations, and the derivative term responds to increases in the violation signal. The resulting multiplier adaptively adjusts the policy optimization pressure according to the evolution of the barrier constraint. This design directly connects policy improvement to the barrier conditions subsequently used in our certification procedure.

Our existing experiments provide evidence for the contribution of PID feedback within barrier-guided learning. Specifically, as described in Section 4.1, Lag-Barrier uses an integral-only multiplier update KP=KD=0K_P=K_D=0 while retaining the same barrier-based feedback. Thus, the comparison between Lag-Barrier and PID-Barrier allows us to examine the benefit of incorporating proportional and derivative feedback. On the Van der Pol invariant task, PID-Barrier achieves a higher mean return (189.79 versus 128.23) and a lower mean control cost (0.024 versus 0.230), while obtaining valid certificates in all five runs, compared with four out of five for Lag-Barrier.

Furthermore, Appendix I (Figs. 6--9) examines the effects of the individual PID components through comparisons of the P, PI, and PID-Barrier variants in terms of learning curves and certification slacks. The results show that the benefits of PID feedback are particularly evident on the Van der Pol system, whose nonlinear dynamics induce substantial trajectory fluctuations and make policy updates more susceptible to short-term disturbances. PID-Barrier exhibits more favorable return–cost behavior and maintains stable verification results within the prescribed bounds.

We acknowledge that a matched PID-Lagrangian baseline using task-level safety costs would provide a complementary comparison to isolate the effect of the feedback signal under the same PID update mechanism. However, our current contribution is specifically the integration of PID feedback with barrier-based policy learning, rather than a claim that barrier-based feedback universally outperforms cost-based feedback under all PID-Lagrangian formulations. We will clarify this distinction and the scope of our experimental conclusions in the revised manuscript.

AuthorsOpenReview ↗

Response 4

Weakness 5: Practical meaning of the PAC guarantee

We thank the reviewer for this insightful question. Our work considers black-box deterministic systems whose transition dynamics are unknown and cannot be directly used for model-based verification. Although the system can be queried through simulation or execution from selected initial states, finite trajectory observations cannot generally establish whether an arbitrary, previously unqueried initial state satisfies the safety specification. Moreover, for a fixed initial state and controller, satisfaction of the safety specification is a deterministic property rather than a probabilistic one. Accordingly, our PAC guarantee is formulated at the set level: it quantifies, with high confidence over the sampled certification data, the proportion of initial states within a certified set that satisfy the corresponding finite-horizon specification, with respect to the prescribed reference measure.

The practical meaning of reuse is that a single offline certification result can support repeated use of the same certified operating region without requiring repeated trajectory sampling or scenario optimization. Specifically, we collect trajectory data from the black-box system and perform scenario optimization offline to obtain the certified set. Once the controller, system dynamics, barrier function, and certified set are fixed, membership of a queried state can be checked by evaluating the neural barrier function and comparing its value with the prescribed threshold, without requiring further interaction with the black-box system.

Importantly, this reuse does not imply that the membership test establishes the safety of each individual state. Rather, the original PAC guarantee remains attached to the certified set as a whole: with the stated confidence, a sufficiently large proportion of initial states in this set satisfy the specified finite-horizon property with respect to the prescribed reference measure. Thus, the guarantee is set-level rather than pointwise. We will clarify the black-box setting, the set-level nature of the PAC guarantee, and the practical meaning of certificate reuse in the revised manuscript.

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

The paper combines neural barrier learning with post-training PAC certification for black-box systems. The result is largely an application of existing scenario-optimization ideas rather than a new certification principle. The resulting PAC guarantee is only set-level: it bounds the fraction of states in a certified region that may violate the property, not whether a particular current state is safe. I'd say the work feels several years late in the direction -- compared to Xue (2019, 2020), Mathiesen (2022) the paper does not strengthen the guarantees much or address a fundamentally harder certification setting.

Critical strengths
  • Extends PAC guarantees to multi-step invariance and reach-avoidance
  • One certification dataset supports multiple certified-set thresholds
Critical weaknesses
  • The PAC guarantee is at the set-level and it's hard to say how that applies to state-specific safety questions.
  • The statistical bound largely inherits standard scenario-optimization guarantees rather than capturing neural-network or system complexity.
  • The method does not clearly improve the strength or scalability of earlier PAC/neural-barrier certification work.
  • Experiments are limited to small classical systems.
Questions for authors
  • What practical safety claim can be made for a specific state encountered online, given that the certificate only bounds the measure of violating states in the set?
  • Beyond applying scenario optimization after freezing the neural network, what is technically new in the certification result relative to earlier PAC safety and reachability work?
  • Can the method compare with stronger or more modern statistical certification approaches beyond simple Monte Carlo baseline?
3 comments
AuthorsOpenReview ↗

Response 1

Weakness 1 and Question 1: Set-level guarantees and state-specific safety.

We thank the reviewer for this insightful question. The difficulty lies in determining whether a previously unqueried state is safe from finite data. Without additional structural information, such as a known Lipschitz bound, observations from other states generally cannot settle this question. Moreover, for the deterministic system considered in our paper, a fixed state under a fixed controller either satisfies the safety specification or violates it. Its safety is therefore a deterministic property, not a probabilistic one. Accordingly, our PAC guarantee is formulated at the set level: it quantifies the proportion of states within a certified set that satisfy the specification.

This motivates our PAC formulation. The reference measure μS\mu_S is a fixed, time-independent measure on the state space, introduced solely to quantify the size of subsets of SS. The PAC certificate is constructed only once using i.i.d. samples drawn from μS\mu_S. With confidence at least 1−β1-\beta, it bounds the measure of states in S~ζ\widetilde{S}_{\zeta} whose kmax⁡k_{\max}-step evolution violates the specification by α\alpha. Consequently, at least a fraction 1−α/μS(S~ζ)1-\alpha/\mu_S(\widetilde{S}_{\zeta}) of a positive-measure certified set satisfies the specification. This guarantee is a property of the certified set itself and is independent of any particular trajectory realization.

The practical benefit is that the learned controller comes with an explicitly characterized region and a quantitative bound on its violating portion. During deployment, membership in this region can be checked using the learned barrier, and the same certificate remains applicable while the controller and dynamics are fixed. Its interpretation continues to be the original finite-horizon statement under μS\mu_S. We will make this operational meaning explicit in the revised manuscript.

Weakness 2: The statistical bound largely inherits standard scenario-optimization guarantees rather than capturing neural-network or system complexity.

We thank the reviewer for raising this point. Avoiding the verification difficulty induced by neural-network nonconvexity and the complexity of unknown dynamics is a central advantage of our framework. Neural networks offer expressive representations for both the controller and the barrier function. However, incorporating their parameters directly into a scenario optimization formulation generally results in a nonconvex optimization problem that is computationally challenging to solve. Obtaining useful guarantees in that setting can require additional work, such as constructing a small support subsample, also called a compression set. Finding such a set can be difficult and time-consuming because it may require repeatedly solving the underlying nonconvex problem.

Our framework avoids this difficulty by using the learned network through its evaluations on sampled system trajectories, thereby enabling a formal characterization of their performance. Once the policy and barrier are fixed, these evaluations determine the coefficients of the certification constraints, and only the scalar slack is optimized. This makes the certification problem convex regardless of the nonconvexity of the trained network or the unknown dynamics. The same scalar scenario argument therefore applies across different neural representations and systems under our assumptions.

The neural barrier plays a central role throughout this construction. Its values determine the trajectory inequalities, the resulting slack, and the shape and size of the certified sets. Our framework therefore exploits the expressive representation learned by the neural network while avoiding the need to solve a global nonconvex verification problem or search for a compression set. We will make this motivation and advantage more explicit in the revised manuscript.

AuthorsOpenReview ↗

Response 2

Weaknesses 3 and 4: Scalability and experimental evaluation.

We thank the reviewer for raising these points. A key scalability advantage of our framework is that its PAC certification sample bound does not grow with the state dimension or neural-network size. In scenario-based certificate synthesis, directly optimizing the certificate parameters can lead to sample bounds that grow with the number of decision variables, making richer certificate representations more data-demanding to certify (Campi and Garatti, 2008). Our framework avoids this dependence by separating neural training from certification: after training, we fix the controller and barrier, and our certification LP optimizes a single scalar ξ\xi.

Specifically, for a prescribed violation-measure bound α\alpha and confidence level 1−β1-\beta, our sufficient trajectory sample size is N≥2α(ln⁡1β+1).N\ge\frac{2}{\alpha}\left(\ln\frac{1}{\beta}+1\right). This sample requirement depends only on α\alpha and β\beta, but not on the state dimension or the number of neural-network parameters. Consequently, increasing the system dimension or using a larger neural network does not itself increase the number of trajectories required by our PAC bound. This allows us to use expressive neural barriers while retaining a sample-efficient certification procedure, providing a theoretical basis for the scalability of our method.

We agree that additional experiments would demonstrate this advantage more clearly. We will add experiments on additional high-dimensional systems in the final version to better demonstrate the sample efficiency and scalability of the proposed method.

Question 2. Beyond applying scenario optimization after freezing the neural network, what is technically new relative to earlier PAC safety and reachability work?

We thank the reviewer for this important question. Our contribution is a certification framework that turns a learned policy and neural barrier into a family of PAC multi-step safe invariant and reach-avoid sets for black-box systems. This addresses a practical gap in neural barrier learning, since minimizing barrier violations on training data alone leaves the behavior on unseen initial states uncertified. Our framework closes this gap using independent trajectory data while retaining the flexibility of neural policy and barrier representations.

Relative to (Wu et al.(2026)), we extend the certification objective from the existence of a suitable input for one-step controlled invariance to the multi-step behavior of the actual learned controller, including safe target reaching. This requires relating trajectory-level barrier violations to the desired specifications. In particular, Theorems1 and2 establish how a calibrated slack determines the barrier level sets for which the corresponding finite-horizon property holds, up to a bounded measure of exceptional states. This construction also accommodates positive slack in the admissible range, allowing useful certificates even when the original barrier inequalities are imperfectly satisfied.

A further distinction lies in the underlying learning and certification paradigms. Wu et al. (2026) adopts a uniform-learning approach, whereas our work follows a learning-then-verification paradigm. Specifically, we first learn the policy and neural barrier and then use independent trajectory samples to statistically certify the multi-step safety and reach-avoid properties of the resulting learned controller. This separation allows us to establish PAC guarantees for the actual learned networks without requiring uniform guarantees over the entire function class. As a result, unlike Wu et al. (2026), our framework does not impose a convexity requirement on the learned functions used to characterize invariant and reach-avoid sets. It accommodates general, potentially nonconvex neural barrier functions, offering greater flexibility in representing the geometry of certified sets.

While our statistical guarantee builds on an established scenario theorem, our methodological contribution lies in the trajectory-based certification formulation and its application to constructing a family of PAC-certified multi-step operating sets. In particular, for a fixed learned policy, neural barrier, and time horizon, a single certification procedure supports all admissible barrier thresholds, allowing users to adjust the certified operating region without retraining the networks or collecting additional certification data. We will clarify these contributions and their distinctions from prior work in the revised manuscript.

AuthorsOpenReview ↗

Response 3

Question 3. Can the method compare with stronger or more modern statistical certification approaches beyond a simple Monte Carlo baseline?

We thank the reviewer for this insightful suggestion. We agree that comparing our framework only with a simple Monte Carlo baseline does not sufficiently demonstrate its advantages relative to established statistical certification methods. We will therefore include an exact one-sided binomial certification baseline based on the Clopper–Pearson confidence bound.

For a fixed candidate set selected independently of the validation data, the exact one-sided Clopper-Pearson confidence bound provides a natural baseline. Suppose that kk violations are observed among NN independent validation trajectory. The corresponding upper confidence bound at confidence level 1−β1-\beta is Beta⁡−1(1−β;k+1,N−k).\operatorname{Beta}^{-1}(1-\beta;k+1,N-k). In the zero-failure regime k=0k=0, this reduces to 1−β1/N.1-\beta^{1/N}. Consequently, certifying that the violation probability is at most α\alpha requires N≥ln⁡(1/β)−ln⁡(1−α)=O(1/α).N\geq \frac{\ln(1/\beta)}{-\ln(1-\alpha)}=O(1/\alpha). for a fixed confidence level. Thus, we acknowledge that our O(1/α)O(1/\alpha) sample-complexity dependence does not constitute an asymptotic improvement over exact binomial certification in the zero-failure regime. However, when violations are observed, the required sample size depends on their frequency and the desired accuracy. In the general nonzero-failure regime, where the violation probability must be estimated to an additive accuracy of order α\alpha, the usual concentration-based sample complexity is O(1/α2)O(1/\alpha^2) for fixed confidence. The exact binomial bound can have different scaling in rare-event regimes, depending on the number of observed violations.

However, the comparison above concerns a fixed certification event and does not capture the main advantage of our framework, which lies in its treatment of data-dependent set selection. Specifically, our calibrated slack ξ∗\xi^\ast is obtained by solving a scenario optimization problem using the certification data. The scenario theorem then provides an out-of-sample probabilistic guarantee for this data-dependent solution, accounting for its dependence on the sampled trajectories. Under the assumptions of the scenario theorem, our resulting sample-complexity bound retains an O(1/α)O(1/\alpha) dependence. Furthermore, Theorems1 and2 establish a deterministic connection between the calibrated slack and a family of barrier level sets. Consequently, a single certification procedure supports PAC guarantees for all admissible barrier thresholds, without requiring separate trajectory validation for each candidate set.

In contrast, applying an ordinary Clopper-Pearson bound to a candidate set selected using the same validation data does not automatically preserve its nominal confidence guarantee. Independent validation or an appropriate simultaneous inference procedure is needed to account for data-dependent selection. Our framework therefore provides an integrated approach to selection-aware PAC certification of a family of multi-step operating sets.

Finally, we emphasize that the primary contribution of our work is not a new statistical confidence bound, but a framework for PAC certification of multi-step safe invariant and reach-avoid sets for black-box systems under learned controllers. To the best of our knowledge, this is the first work to establish such guarantees within a learning-then-verification framework using trajectory-based scenario optimization. We will clarify these distinctions and include the Clopper-Pearson baseline in the revised manuscript to better demonstrate the statistical contributions of our framework.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.