Learning Quasi-LPV Models and Robust Control Invariant Sets with Reduced Conservativeness

Alberto Bemporad; Sampath Kumar Mulagaleti

REVIEW 3 major objections 6 minor 25 references

Learning Quasi-LPV Models and Robust Control Invariant Sets with Reduced Conservativeness

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read The paper claims that identifying a qLPV model together with a robust control invariant set can be made less conservative by shrinking the model's multiplicative uncertainty with interval bound propagation on the softmax scheduling…

desk verdict A correct and practically useful tightening of the authors' own concurrent qLPV identification scheme, but the advertised reduction in conservativeness is verified only for one simulation and rests on an unverified disturbance-set inclusion. read the letter →

arxiv 2505.07287 v1 pith:YOVCBYQS submitted 2025-05-12 math.OC cs.SYeess.SY

Sampath Kumar Mulagaleti , Alberto Bemporad This is my paper · ORCID

classification math.OCcs.SYeess.SY MSC 93B3093D0990C4768T07

keywords quasi-LPVmodelsrobustcontrolinvariantsetsconcurrentsynthesiscontrol-orientedregularizationintervalboundpropagationmultiplicativeuncertaintysoftmaxschedulingsystemidentification

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that a quasi-linear parameter-varying (qLPV) model learned from input-output data can be certified, at the same time as it is identified, to admit a robust control invariant (RCI) set, meaning a set of states from which a controller can keep the plant inside given output and input constraints despite disturbances, and that the certified set can be considerably less conservative than in previous work. The mechanism is a tightened uncertain linear model: instead of letting the scheduling vector range over the whole simplex, the paper uses lower bounds on each softmax scheduling function over the candidate invariant set to shift the vertex matrices and shrink the multiplicative uncertainty. The lower bounds are computed by interval bound propagation through the scheduling network, so the whole pipeline remains differentiable and can be embedded in gradient-based system identification. If the claim is right, concurrent identification of models and safety guarantees becomes less conservative without degrading prediction accuracy.

What carries the argument

The load-bearing object is the tightened multiplicative-uncertainty hull $\tilde\Delta \subseteq \Delta$ in (21)-(22), together with the interval-bound-propagation step that produces valid lower bounds $a_i \le p_i(z)$ on an inflated bounding box $B(q)$ around the candidate invariant set. Proposition 3 shows that invariance with respect to $\tilde\Delta$ is sufficient for invariance of the qLPV model, and Corollary 1 turns that into the constraint set $\tilde S$. Algorithm 1 alternates bound propagation (step 2) with a differentiable quadratic program over $\hat S(q_k,a)$ (step 3), so the regularization value is differentiable and can serve as a term in the identification objective. The configuration-constrained polytope parameterization $X(q)=\mathrm{CH}\{V_j q,\ j\in I_v^1\}$ is what allows the RCI conditions to be written as finite linear inequalities.

What would settle it

Collect a long validation set from the true plant, simulate the observer (5) with the identified model, and form the samples $w_t=y_t-Cz_t^w$; if any sample lies outside the box $\mathcal W=\{w: |w-c_w|\le 1.01\,\epsilon_w\}$ used at identification time, the RCI certificate is invalid for the plant. Equivalently, run the closed-loop controller (27) beyond the training horizon; a single trajectory leaving $Y$ disproves the guarantee.

Watch

Extended reading notes

Core claim

The central claim is Proposition 3: if, on the candidate configuration-constrained polytope $X(q)$, each scheduling component satisfies $p_i(z)\ge a_i$, then the qLPV system is contained in an uncertain LTI system $\tilde\Delta$ whose vertices are the shifted matrices $\tilde A_i=(1-\sum_{j=1}^{n_p} a_j)A_i+\sum_{j=1}^{n_p} a_j A_j$ (and similarly for $B$ and $L$). Consequently any robust control invariant set certified for $\tilde\Delta$ is automatically invariant for the qLPV model. The paper defines the new regularization $r(\Theta)=\inf_{(q,v)\in\tilde S} d(\bar A,\bar B,C,q)$, where $\tilde S$ is the constraint set from Corollary 1; since $\tilde S$ contains the earlier set $S$ (recovered by setting $a=0$), the new value is always at least as good and is typically strictly better. Algorithm 1 evaluates this regularization by alternating interval bound propagation with a differentiable quadratic program, which is what allows it to be used inside the identification loop. On the oscillator example the identified models achieve $d$-values significantly below both $d_{\mathrm{seq}}=22.6924$ for sequential identification and $d_{\mathrm{base}}=5.0936$ for the prior concurrent approach, with best-fit rates near 90.7 percent.

Load-bearing premise

The data-driven disturbance box $\mathcal W$, inflated by $\kappa=1.01$, is assumed to contain all disturbances of the true plant; if it does not, the certified invariant set says nothing about the physical system.

Editorial extensions

If this is right

Any qLPV model returned by Algorithm 2 carries the certificate that an RCI set exists, so a tracking controller that keeps the plant output inside $Y$ while respecting input limits can be synthesized from the identified model.
Because $S\subseteq\tilde S$, the improved regularization never produces a worse RCI-size score than the earlier QP-based regularization on the same model, and it yields strictly better scores whenever the scheduling lower bounds are informative.
On the oscillator test problem, the approach achieves $d$-values below both benchmarks (22.6924 and 5.0936) while the best-fit rate stays in the narrow band 90.73-90.74 on the training data, so the conservativeness reduction does not come at the cost of model quality.
Running Algorithm 1 for one iteration inside the identification loop is enough for the reported gains, and the value $r_k$ is observed to converge monotonically in the post-hoc evaluation of Figure 2.
Because the proof of Proposition 3 only uses $p(z)\in\mathcal P$, the same tightening applies when the scheduling function also depends on the input, as pointed out in Remark 1.

Reading between the lines

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

Editorial inference: the same lower-bound tightening could be applied with exact polytope bounds on the scheduling network instead of an inflated box, which may produce even tighter $\tilde\Delta$ and smaller $d$-values at higher computational cost.
Editorial inference: the comparison between $\zeta$ values is metric-dependent; a metric based on the volume of $X(q)$ rather than the tracking-distance $d$ could rank the identified models differently, so the reported gain should be read as specific to the tracking formulation.
Editorial inference: a natural testable extension is to make $\zeta$ adaptive during training, shrinking the bounding box as the model improves, which could remove the need to grid over $\zeta$ and improve convergence behavior.
Editorial inference: the approach's reliance on monotone activations is a practical limitation; replacing interval bound propagation with a differentiable verifier for general activations is the obvious next step and is flagged by the authors as future work.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The core step here is Proposition 3, and it is correct: by using per-state lower bounds on the softmax scheduling functions, the qLPV model is enclosed in a tighter multiplicative uncertainty set, and the set-inclusion proof is clean. The observation that a = 0 recovers the earlier QP formulation means the new condition is no more conservative than the old one—that is exactly the right kind of extension. Building the new regularization on interval bound propagation is also a legitimate and practical refinement, and the released code plus the differentiability argument make it usable.

The numerical demonstration does show the claimed effect: for the oscillator example, the d-values fall well below both d_seq and d_base while the BFR stays essentially flat. So the method appears to work on the test problem.

The soft spots are real but not fatal. The most important is the disturbance set W in (12). The paper cites [1, Prop. 4] for a lower bound on κ that would certify W = CE ⊕ V, then simply sets κ = 1.01 in the experiments without stating or checking that bound. If the required κ is larger, then the RCI sets certified for the uncertain model are not certified for the physical oscillator. This makes the central claim conditional, and the paper itself only says 'sufficiently large κ' in the controller-synthesis section without quantifying it. That needs to be fixed.

Second, Algorithm 1 has no convergence proof. The authors acknowledge this and show one monotonic run in Figure 2, but a single trajectory is not evidence of general behavior. Third, the numerical study is one plant, one random dataset, no seed variation, and the comparison with d_seq is somewhat apples-to-oranges because the sequential benchmark uses a maximal RCI set while the concurrent benchmarks use configuration-constrained polytopes. The comparison with d_base is fair, and that alone carries the point.

The reliance on the authors' prior work is heavy but honest—[1] contains the base architecture, the W characterization, and the initialization. Nothing here is hidden.

This paper is for researchers working on LPV identification with formal control guarantees. It deserves a serious referee, not a desk reject. I would recommend acceptance after the authors verify the κ condition on their example, add at least a seed variation or a second example, and either prove or more carefully qualify the convergence of Algorithm 1.

Referee Report

3 major / 6 minor

Summary. The paper proposes a concurrent identification framework for qLPV models and robust control invariant (RCI) sets, extending the authors' earlier work [1]. The main idea is to replace the full simplex-valued uncertainty set of the scheduling-dependent model with a tightened convex hull \tilde\Delta, obtained by using lower bounds a_i on each scheduling function p_i(z) over the candidate RCI set. This yields a new regularization function r(\Theta) defined as the optimal value of a nonlinear robust optimization problem, which is then evaluated by an iterative algorithm that alternates interval bound propagation (IBP) with QP solves. The resulting differentiable surrogate is embedded in a gradient-based training loop for joint model identification and RCI-set construction. A numerical oscillator example reports that the obtained RCI sets have markedly smaller size values d_\zeta than two benchmarks, while BFR scores remain around 90.7-90.9.

Significance. The tightening idea is sound and potentially useful: Proposition 3 and Corollary 1 give a clean construction in which a lower-bound vector a on the scheduling functions over X(q) yields a less conservative uncertain LTI enclosure, and the observation that a=0 recovers the previous formulation guarantees no additional conservativeness in that branch. The reported numerical reductions (d_\zeta significantly below d_seq=22.6924 and d_base=5.0936) are encouraging, and the release of code is a positive feature. The strength of the paper is the derivation itself; its main weaknesses are that the plant-level guarantee depends on an unverified disturbance-set inclusion and that the algorithm used to evaluate the central regularization function has no stated convergence or approximation guarantees.

major comments (3)

[Sec. 5.0.1, Eq. (12) and Proposition 1] The numerical demonstration of the central claim is conditional on an unverified disturbance-set inclusion. Proposition 1 requires W = CE \oplus V in (7); Section 2.2.1 defines W in (12) as a \kappa-inflated box and refers to [1, Prop. 4] for lower bounds on \kappa. Section 5.0.1 fixes \kappa = 1.01 without reporting whether that value satisfies the required lower bound for the identified model. If the true required \kappa exceeds 1.01, the RCI sets certified for (21) and the d_\zeta values in Figure 1 are not certified for the oscillator plant, and the closed-loop trajectories in Figure 3 lack the claimed guarantee. The statement in Section 5.0.2 that 'For sufficiently large values of \kappa>1' Problem (27) is recursively feasible only reinforces the need to verify or quantify this assumption. Please verify the lower bound for the identified model or explicitly restrict the numerical claims to the auxiliary uncertain model.
[Sec. 3.2 and Algorithm 1] The paper replaces the nonlinear robust optimization problem (23) by a finite sequence of QPs over \hat S(q_k,a_k) and uses the output r_{\hat k} as the regularization value, but no convergence, monotonicity, or even feasibility guarantee is provided for Algorithm 1; the text states that theoretical properties are future research. Consequently, the d_\zeta numbers in Figure 1 are upper bounds that depend on the initialization q_0, on \zeta, and on \hat k (here \hat k=200), rather than solutions of (23). Since r(\Theta;q_0) is then used as a differentiable objective in Algorithm 2, the training procedure optimizes a surrogate whose relationship to the exact r(\Theta) is uncharacterized. Please provide at least a feasibility/monotonicity statement for Algorithm 1 or report the results as feasible RCI-set sizes rather than as values of the regularizer.
[Corollary 1, definition of \tilde S] The definition of \tilde S is formally ill-posed. It reads '\exists a \in [0,p(z)], \forall z\in X(q)' inside the set description, which if interpreted literally allows a to depend on z. However, Proposition 3 and the proof of Corollary 1 require a single vector a satisfying 0 \le a_i \le p_i(z) for all z\in X(q). Please restate the existential quantifier outside the universal quantification over z, i.e., 'there exists a such that 0\le a_i\le p_i(z) for all z\in X(q)', before the vertex constraints.

minor comments (6)

[Proposition 3 proof] The proof divides by 1 - \sum_j a_j, which vanishes when \sum_j a_j = 1; this case is not treated separately. The conclusion is still true (in that case p(z)=a on X(q) and \tilde\Delta is a singleton), but the proof should handle it explicitly or by a continuity argument.
[Corollary 1 proof] In Proof 2, the phrase 'observing that S is RCI for (21)' should read 'observing that X(q) is RCI for (21)'.
[Sec. 2.2.1, Eq. (12)] The componentwise nature of \bar w := \max_{w\in W} w and \underline w := \min_{w\in W} w should be stated explicitly, since W is a set of vectors.
[Figure 1 caption] The caption does not identify the dashed horizontal lines; please label the d_seq and d_base values and distinguish them from the plotted d_\zeta curve.
[Sec. 5.0.1] The claim that choosing \hat k>1 'often results in Algorithm 2 converging to suboptimal points' is unsupported; either provide evidence or soften this statement.
[Reference list] Reference [19] contains a typographical error in the journal name: 'IIEEE' should be 'IEEE'.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; the derivation is self-contained. The main risk is an unverified κ assumption, which is a correctness issue rather than a circular step.

full rationale

The paper's central claim—that the proposed control-oriented regularization yields qLPV models admitting RCI sets with reduced conservativeness—is supported by an independent derivation. Proposition 3 provides a new mathematical condition (a lower-bound-dependent tightening of the multiplicative uncertainty set), and its proof is a self-contained convex-hull argument. Proposition 4 computes the needed lower bounds via interval bound propagation from the neural network scheduling functions, without presupposing the target. Algorithm 1 evaluates the regularization by alternating between bound propagation and QP optimization, and the final RCI sets are certified on the constructed sets. The numerical comparison uses the same size metric d(·) for all methods, and the improvement is not a construction artifact: the paper reports no reduction when Algorithm 1 is applied to baseline models, indicating the benefit arises from training the model to have favorable lower bounds. The reliance on the authors' prior work [1] is for the concurrent-synthesis framework, the W characterization, and the baseline benchmark; these are background and comparison, not the novel claim. The main weakness is that κ=1.01 in (12) is assumed (citing [1, Prop. 4]) to satisfy the lower bound guaranteeing W=CE⊕V, without verifying it for the identified model; if this fails, the plant-level guarantee does not hold. This is an unverified correctness condition, not a circular definition. No step in the derivation reduces to its own input by construction.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the qLPV model class, the inherited data-driven disturbance characterization (with user-chosen kappa), the soundness of IBP for monotonic networks, and the configuration-constrained polytope machinery. No new physical entities are introduced; the new items are algorithmic and depend on the listed parameters and assumptions.

free parameters (8)

qLPV scheduling network weights theta_i = not reported (trained on D)
The softmax scheduling functions are feedforward networks; their weights are optimized in (10) and directly determine the tightened uncertainty set.
model matrices A_i, B_i, C, L_i = not reported (estimated from D and D_w)
Identified parameters of (2) and observer gains in (5); central to the RCI regularization.
initial state x0 = not reported
Optimized in (10) as part of identification.
kappa, disturbance inflation = 1.01
User-specified in Section 5.0.1; assumed to verify (7) via [1, Prop. 4].
zeta, bounding-box inflation = 0.01 to 0.1
User-specified sweep in Section 5; controls tightness of B(q) in (24).
tau, regularization weight = 0.0005
Weights prediction loss against RCI-size regularization in (10).
M, tracking horizon in size measure = 5
Defines the RCI-size proxy d in (17).
architecture sizes nx, np, hidden units, f, v = nx=2, np=6, hidden=3, f=v=4
Hyperparameters chosen in Section 5.0.1; they shape the model class and the configuration-constrained polytope.

assumptions (5)

domain assumption The true plant can be represented by (6) with an output error bound V and an observer invariant set E satisfying hat x0 - z0 in E implying hat xt - zt in E for all t.
Proposition 1 inherits this from [1]; it is not verified on the plant, only assumed for the safety argument.
domain assumption The disturbance set W in (12) with kappa = 1.01 contains all plant-model mismatch disturbances, i.e., W = C E direct sum V with E and V as above.
Section 2.2.1 relies on [1, Prop. 4] for lower bounds on kappa; no data-driven certificate is computed in this paper.
domain assumption FNN activations are monotonic, so interval bound propagation through the softmax composition yields valid componentwise bounds (26).
Section 3.3; the example uses elu(x)+1, which is monotonic; nonmonotonic activations are explicitly deferred to future work.
standard math Configuration-constrained polytope representation X(q) = CH{Vj q, j in I^v_1} and vertex maps Vj from [13] are available for the chosen F.
Section 2.2.2 imports this prior result from [13].
domain assumption The differentiable QP solver qpax computes gradients of Algorithm 1 sufficiently accurately for gradient-based training.
Section 5 uses qpax [23]; numerical behavior of the solver is not analyzed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Learning Quasi-LPV Models and Robust Control Invariant Sets with Reduced Conservativeness." pith.science (2026). https://pith.science/paper/YOVCBYQS

@misc{pith2026250507287,
  author       = {Pith},
  title        = {Pith review of: Learning Quasi-LPV Models and Robust Control Invariant Sets with Reduced Conservativeness},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YOVCBYQS}},
  note         = {Machine review of arXiv:2505.07287}
}

read the original abstract

We present an approach to identify a quasi Linear Parameter Varying (qLPV) model of a plant, with the qLPV model guaranteed to admit a robust control invariant (RCI) set. It builds upon the concurrent synthesis framework presented in [1], in which the requirement of existence of an RCI set is modeled as a control-oriented regularization. Here, we reduce the conservativeness of the approach by bounding the qLPV system with an uncertain LTI system, which we derive using bound propagation approaches. The resulting regularization function is the optimal value of a nonlinear robust optimization problem that we solve via a differentiable algorithm. We numerically demonstrate the benefits of the proposed approach over two benchmark approaches.

Figures

Figures reproduced from arXiv: 2505.07287 by the authors.

**Figure 1.** Output of Algorithm 1 for different ζ, simulated using models Θζ identified by Algorithm 2. Observe that for chosen ζ, we obtain dζ significantly lesser than dseq = 22.6924 and dbase = 5.0936 [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗

**Figure 2.** Iterations of Algorithm 1 for Θ = Θ0.07 and ζ = 0.07. To compare against these benchmarks, we simulate Algorithm 2 with ζ uniformly spaced in [0.01, 0.1], and fix ˆk = 1 such that we perform one iteration of Algorithm 1 per iteration of Algorithm 2. We report that while ˆk > 1 can be chosen to simulate Algorithm 1, it often results in Algorithm 2 converging to suboptimal points. A study of escaping such minima is a … view at source ↗

**Figure 3.** Closed-loop trajectories using tracking controller (27). The black region in top figure denotes boundaries of Y [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 21 canonical work pages

[1]

Combined learning of linear parameter-varying models and robust control invariant sets,

S. K. Mulagaleti and A. Bemporad, “Combined learning of linear parameter-varying models and robust control invariant sets,” arXiv preprint arXiv:2411.18166, 2024

work page arXiv 2024
[2]

Deep learning and system identification,

L. Ljung, C. Andersson, K. Tiels, and T. B. Sch¨ on, “Deep learning and system identification,” IFAC-PapersOnLine, vol. 53, no. 2, pp. 1175–1181, 2020. 21st IFAC World Congress

work page 2020
[3]

Set-membership identification of systems with parametric and nonparametric uncertainty,

R. L. Kosut, M. K. Lau, and S. P. Boyd, “Set-membership identification of systems with parametric and nonparametric uncertainty,” IEEE Trans. Autom. Control, vol. 37, no. 7, pp. 929–941, 1992

work page 1992
[4]

Set membership identification of nonlinear systems,

M. Milanese and C. Novara, “Set membership identification of nonlinear systems,” Automatica, vol. 40, no. 6, pp. 957–975, 2004

work page 2004
[5]

Set membership identification of linear systems with guaranteed simu- lation accuracy,

M. Lauricella and L. Fagiano, “Set membership identification of linear systems with guaranteed simu- lation accuracy,” IEEE Trans. Autom. Control, vol. 65, no. 12, pp. 5189–5204, 2020

work page 2020
[6]

Data-driven computation of minimal robust control invariant set,

Y. Chen, H. Peng, J. Grizzle, and N. Ozay, “Data-driven computation of minimal robust control invariant set,” in 2018 IEEE Conference on Decision and Control (CDC), pp. 4052–4058, 2018

work page 2018
[7]

Data-driven synthesis of robust invariant sets and controllers,

S. K. Mulagaleti, A. Bemporad, and M. Zanon, “Data-driven synthesis of robust invariant sets and controllers,” IEEE Control Systems Letters, vol. 6, pp. 1676–1681, 2022

work page 2022
[8]

Safe reinforcement learning using robust MPC,

M. Zanon and S. Gros, “Safe reinforcement learning using robust MPC,” IEEE Trans. Autom. Control, vol. 66, no. 8, pp. 3638–3652, 2021

work page 2021

Show all 25 references

[9]

Lovera, M

M. Lovera, M. Bergamasco, and F. Casella, LPV Modelling and Identification: An Overview. Springer Berlin Heidelberg, 2013

2013
[10]

State-space LPV model identification using kernelized machine learning,

S. Z. Rizvi, J. M. Velni, F. Abbasi, R. T´ oth, and N. Meskin, “State-space LPV model identification using kernelized machine learning,” Automatica, vol. 88, pp. 38–47, 2018

2018
[11]

Deep-learning-based identification of LPV models for nonlinear systems,

C. Verhoek, G. I. Beintema, S. Haesaert, M. Schoukens, and R. T´ oth, “Deep-learning-based identification of LPV models for nonlinear systems,” in 2022 IEEE 61st Conference on Decision and Control (CDC), pp. 3274–3280, 2022

2022
[12]

Blanchini and S

F. Blanchini and S. Miani, Set-Theoretic Methods in Control. Springer International Publishing, 2015

2015
[13]

Configuration-constrained tube MPC,

M. E. Villanueva, M. A. M¨ uller, and B. Houska, “Configuration-constrained tube MPC,” Automatica, vol. 163, p. 111543, 2024

2024
[14]

Robust output feedback model predictive control of constrained linear systems,

D. Mayne, S. Rakovi´ c, R. Findeisen, and F. Allg¨ ower, “Robust output feedback model predictive control of constrained linear systems,” Automatica, vol. 42, no. 7, pp. 1217–1222, 2006

2006
[15]

A one-step approach to computing a polytopic robust positively invariant set,

P. Trodden, “A one-step approach to computing a polytopic robust positively invariant set,” IEEE Trans. Autom. Control, vol. 61, no. 12, pp. 4100–4105, 2016

2016
[16]

Configuration- constrained tube MPC for tracking,

F. Badalamenti, S. K. Mulagaleti, A. Bemporad, B. Houska, and M. E. Villanueva, “Configuration- constrained tube MPC for tracking,” IEEE Control Systems Letters, vol. 8, pp. 1174–1179, 2024

2024
[17]

On the effectiveness of interval bound propagation for training verifiably robust models,

S. Gowal, K. Dvijotham, R. Stanforth, R. Bunel, C. Qin, J. Uesato, R. Arandjelovic, T. Mann, and P. Kohli, “On the effectiveness of interval bound propagation for training verifiably robust models,” arXiv preprint arXiv:1810.12715, 2018

arXiv
[18]

Beta-crown: Efficient bound propagation with per-neuron split constraints for neural network robustness verification,

S. Wang, H. Zhang, K. Xu, X. Lin, S. Jana, C.-J. Hsieh, and J. Z. Kolter, “Beta-crown: Efficient bound propagation with per-neuron split constraints for neural network robustness verification,” in Advances in Neural Information Processing Systems, vol. 34, pp. 29909–29921, 2021

2021
[19]

Safety verification and robustness analysis of neural networks via quadratic constraints and semidefinite programming,

M. Fazlyab, M. Morari, and G. J. Pappas, “Safety verification and robustness analysis of neural networks via quadratic constraints and semidefinite programming,” IIEEE Trans. Autom. Control, vol. 67, no. 1, pp. 1–15, 2022

2022
[20]

Linsyn: Synthesizing tight linear bounds for arbitrary neural network activation functions,

B. Paulsen and C. Wang, “Linsyn: Synthesizing tight linear bounds for arbitrary neural network activation functions,” in Tools and Algorithms for the Construction and Analysis of Systems: 28th International Conference, TACAS 2022, (Berlin, Heidelberg), p. 357–376, Springer-Ver...

2022
[21]

Adam: A method for stochastic optimization,

D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2017

2017 arXiv
[22]

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
[23]

On the differentiability of the primal-dual interior-point method,

K. Tracy and Z. Manchester, “On the differentiability of the primal-dual interior-point method,” arXiv preprint arXiv:2406.11749, 2024

2024 arXiv
[24]

An L-BFGS-B approach for linear and nonlinear system identification under ℓ1 and group-lasso regularization,

A. Bemporad, “An L-BFGS-B approach for linear and nonlinear system identification under ℓ1 and group-lasso regularization,” IEEE Trans. Autom. Control, pp. 1–8, 2025

2025
[25]

Infinite time reachability of state-space regions by using feedback control,

D. Bertsekas, “Infinite time reachability of state-space regions by using feedback control,” IEEE Trans. Autom. Control, vol. 17, no. 5, pp. 604–613, 1972

1972

Pith tools

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

[1] [1]

Combined learning of linear parameter-varying models and robust control invariant sets,

S. K. Mulagaleti and A. Bemporad, “Combined learning of linear parameter-varying models and robust control invariant sets,” arXiv preprint arXiv:2411.18166, 2024

work page arXiv 2024

[2] [2]

Deep learning and system identification,

L. Ljung, C. Andersson, K. Tiels, and T. B. Sch¨ on, “Deep learning and system identification,” IFAC-PapersOnLine, vol. 53, no. 2, pp. 1175–1181, 2020. 21st IFAC World Congress

work page 2020

[3] [3]

Set-membership identification of systems with parametric and nonparametric uncertainty,

R. L. Kosut, M. K. Lau, and S. P. Boyd, “Set-membership identification of systems with parametric and nonparametric uncertainty,” IEEE Trans. Autom. Control, vol. 37, no. 7, pp. 929–941, 1992

work page 1992

[4] [4]

Set membership identification of nonlinear systems,

M. Milanese and C. Novara, “Set membership identification of nonlinear systems,” Automatica, vol. 40, no. 6, pp. 957–975, 2004

work page 2004

[5] [5]

Set membership identification of linear systems with guaranteed simu- lation accuracy,

M. Lauricella and L. Fagiano, “Set membership identification of linear systems with guaranteed simu- lation accuracy,” IEEE Trans. Autom. Control, vol. 65, no. 12, pp. 5189–5204, 2020

work page 2020

[6] [6]

Data-driven computation of minimal robust control invariant set,

Y. Chen, H. Peng, J. Grizzle, and N. Ozay, “Data-driven computation of minimal robust control invariant set,” in 2018 IEEE Conference on Decision and Control (CDC), pp. 4052–4058, 2018

work page 2018

[7] [7]

Data-driven synthesis of robust invariant sets and controllers,

S. K. Mulagaleti, A. Bemporad, and M. Zanon, “Data-driven synthesis of robust invariant sets and controllers,” IEEE Control Systems Letters, vol. 6, pp. 1676–1681, 2022

work page 2022

[8] [8]

Safe reinforcement learning using robust MPC,

M. Zanon and S. Gros, “Safe reinforcement learning using robust MPC,” IEEE Trans. Autom. Control, vol. 66, no. 8, pp. 3638–3652, 2021

work page 2021