pith. sign in

arxiv: 2504.11125 · v2 · submitted 2025-04-15 · 📡 eess.SY · cs.SY· math.OC

A mixed-integer framework for analyzing neural network-based controllers for piecewise affine systems with bounded disturbances

Dieter Teichrib , Moritz Schulze Darup This is my paper

Pith reviewed 2026-05-22 20:20 UTC · model grok-4.3

classification 📡 eess.SY cs.SYmath.OC
keywords mixed-integer programmingpiecewise affine systemsneural network controllersrobustly positively invariant setsstability certificationhybrid systemsbounded disturbances
0
0 comments X

The pith

A method encodes the closed-loop dynamics of piecewise affine systems under neural network control and bounded disturbances as mixed-integer linear constraints.

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

The paper develops a way to turn the combined behavior of a piecewise affine plant, a neural network controller, and additive disturbances into a set of mixed-integer linear equalities and inequalities. Once encoded this way, the task of finding a robustly positively invariant set reduces to solving a mixed-integer linear program. If such a set exists, it immediately certifies that all future trajectories remain inside a prescribed region and therefore satisfy stability and safety specifications. The same encoding also covers nonlinear plants by treating the difference between the true dynamics and a piecewise affine approximation as an additional bounded disturbance.

Core claim

We present a method for representing the closed-loop dynamics of piecewise affine (PWA) systems with bounded additive disturbances and neural network-based controllers as mixed-integer (MI) linear constraints. We show that such representations enable the computation of robustly positively invariant (RPI) sets for the specified system class by solving MI linear programs. These RPI sets can subsequently be used to certify stability and constraint satisfaction. Furthermore, the approach allows to handle non-linear systems based on suitable PWA approximations and corresponding error bounds, which can be interpreted as the bounded disturbances from above.

What carries the argument

Mixed-integer linear encoding of the closed-loop map that jointly captures the piecewise affine system, the neural network controller, and the bounded disturbance set.

If this is right

  • RPI sets for the closed-loop system can be obtained directly by solving a mixed-integer linear program.
  • Existence of a nonempty RPI set certifies both asymptotic stability and satisfaction of state and input constraints for all time.
  • Nonlinear plants become admissible once they are replaced by a piecewise affine approximation whose approximation error is treated as an additional bounded disturbance.
  • The same mixed-integer representation can be reused for other verification tasks that rely on invariant-set computation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The framework could be paired with existing hybrid-system reachability tools to obtain tighter over-approximations of reachable sets.
  • If efficient big-M reformulations or cutting-plane methods are applied, the approach might scale to networks with several hundred neurons.
  • The disturbance interpretation of approximation error suggests a direct route to robustness margins when the piecewise affine model is obtained from data.

Load-bearing premise

The neural network controller admits an exact mixed-integer linear encoding and the additive disturbances remain inside known bounds that do not depend on the state or input.

What would settle it

A concrete trajectory starting inside the computed RPI set that eventually leaves the set while the disturbance stays inside its declared bound, or a neural network whose input-output map cannot be represented exactly by the proposed mixed-integer constraints.

Figures

Figures reproduced from arXiv: 2504.11125 by Dieter Teichrib, Moritz Schulze Darup.

Figure 2
Figure 2. Figure 2: The green and blue “tubes” illustrate in each case 100 trajectories of the closed-loop system with a random additive disturbance kd(k)k∞ ≤ 0.15 in every time step. The black lines represent trajectories of the nominal system (i.e. d(k) = 0). The small set around the origin is Rmin and the box [−7.3, 8.91] × [−10, 8.52] is the set Fmax. The thick grey lines indicate the regions of the PWA system. B. PWA app… view at source ↗
Figure 1
Figure 1. Figure 1: Illustration of the computation of the set Rmin = R D k (Fmax) in blue starting from the set Fmax in green. The black sets illustrate the shrinking se￾quence of sets R D i (Fmax) with i ∈ {1, . . . k − 1} during the computation of Rmin. The computation terminates with the set R D k (Fmax) = Rmin when (25) holds. Since we have (26) for α = 1 + 9 × 10−7 , Theorem 4 ap￾plies here. This means that all trajecto… view at source ↗
Figure 3
Figure 3. Figure 3: Nonlinear double integrator with an NN-based controller. The grey arrows represent the evolution of the closed-loop system within the set Fmax (green set). The red line represents a trajectory x(0), . . . ,x(k) of the system (29) with (3) starting at x(0) = (5.5 − 1.5)⊤. The small sets around the states x(i) are the reachable sets R D i [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
read the original abstract

We present a method for representing the closed-loop dynamics of piecewise affine (PWA) systems with bounded additive disturbances and neural network-based controllers as mixed-integer (MI) linear constraints. We show that such representations enable the computation of robustly positively invariant (RPI) sets for the specified system class by solving MI linear programs. These RPI sets can subsequently be used to certify stability and constraint satisfaction. Furthermore, the approach allows to handle non-linear systems based on suitable PWA approximations and corresponding error bounds, which can be interpreted as the bounded disturbances from above.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 2 minor

Summary. The manuscript presents a method for representing the closed-loop dynamics of piecewise affine (PWA) systems with bounded additive disturbances and neural network-based controllers as mixed-integer linear constraints. It shows that such representations enable computation of robustly positively invariant (RPI) sets for this system class by solving mixed-integer linear programs (MILPs). These RPI sets can certify stability and constraint satisfaction. The approach also extends to nonlinear systems via suitable PWA approximations whose approximation errors are treated as state-independent bounded disturbances.

Significance. If the encodings are exact, the framework would supply a practical computational route to RPI-set computation for NN-controlled PWA systems, extending standard MILP encodings of ReLU activations and PWA partitions to include additive disturbances. The central claim rests on well-established hybrid-system techniques rather than on novel axioms or invented entities, and the resulting MILP formulations yield falsifiable predictions (feasibility or infeasibility of the invariance constraints). The stress-test concern that the abstract supplies no derivation steps or validation does not land once the full manuscript is examined; the derivations follow directly from existing exact encodings.

minor comments (2)
  1. [Abstract] Abstract: the phrase 'the neural network controller admits an exact mixed-integer linear encoding' should be qualified by the specific activation functions and network depth for which exactness holds, as this is load-bearing for the claim that the closed-loop representation remains exact.
  2. The manuscript would benefit from an explicit statement of the MILP objective and constraint set used to compute the RPI set (e.g., the form of the invariance condition and the binary variables introduced for the NN and PWA regions).

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment of the manuscript, the accurate summary of our contributions, and the recommendation for minor revision. No major comments were raised in the report.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The derivation relies on standard exact MILP encodings for ReLU activations and PWA partitions (with disturbances as additional bounded variables), which are independent of the RPI-set computation that follows. No equation or claim reduces by construction to a fitted parameter, self-defined quantity, or load-bearing self-citation; the central representation is assembled from known hybrid-system techniques and remains falsifiable against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard domain assumptions about piecewise affine dynamics and bounded disturbances together with the technical premise that neural networks admit exact mixed-integer encodings.

axioms (2)
  • domain assumption The plant is piecewise affine and subject to bounded additive disturbances.
    Explicitly stated in the abstract as the system class under consideration.
  • domain assumption Neural network controllers admit an exact representation as mixed-integer linear constraints.
    This is the core modeling step required for the claimed representation.

pith-pipeline@v0.9.0 · 5625 in / 1287 out tokens · 91295 ms · 2026-05-22T20:20:27.831216+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

26 extracted references · 26 canonical work pages · 1 internal anchor

  1. [1]

    Azuma, J

    S. Azuma, J. Imura, and T. Sugie. Lebesgue piecewise affine approximation of nonlinear systems. Nonlinear Analysis: Hybrid Systems , 4(1):92–102, 2010

    work page 2010

  2. [2]

    Bemporad, G

    A. Bemporad, G. Ferrari-Trecate, and M. Morari. Ob- servability and controllability of piecewise affine and hy- brid systems. IEEE Transactions on Automatic Control , 45(10):1864–1876, 2000

    work page 2000

  3. [3]

    Mignone, G

    D. Mignone, G. Ferrari-Trecate, and M. Morari. Stabilit y and stabilization of piecewise affine and hybrid systems: an lmi approach. In Proc. of the 2000 Conference on Decision and Control, volume 1, pages 504–509 vol.1, 2000

    work page 2000

  4. [4]

    Grieder, M

    P . Grieder, M. Kvasnica, M. Baoti´ c, and M. Morari. Sta- bilizing low complexity feedback control of constrained piecewise affine systems. Automatica, 41(10):1683–1694, 2005

    work page 2005

  5. [5]

    Camacho, D.R

    E.F. Camacho, D.R. Ramirez, D. Limon, D. Mu˜ noz de la Pe˜ na, and T. Alamo. Model predictive control techniques for hybrid systems. Annual Reviews in Control , 34(1):21– 31, 2010

    work page 2010

  6. [6]

    Teichrib and M

    D. Teichrib and M. Schulze Darup. Reachability analy- sis for piecewise affine systems with neural network-based controllers. In Proc. of the 2024 Conference on Decision and Control, pages 894–900, 2024

    work page 2024

  7. [7]

    M. S. Ghasemi and A. A. Afzalian. Robust tube-based mpc of constrained piecewise affine systems with bounded ad- ditive disturbances. Nonlinear Analysis: Hybrid Systems , 26:86–100, 2017

    work page 2017

  8. [8]

    D. Q. Mayne, M. M. Seron, and S. V . Rakovi´ c. Robust model predictive control of constrained linear systems with bounded disturbances. Automatica, 41:219–224, 2005

    work page 2005

  9. [9]

    Jones and M

    C.N. Jones and M. Morari. Approximate explicit mpc us- ing bilevel optimization. In Proc. of the 2009 European Control Conference, pages 2396–2401, 2009

    work page 2009

  10. [10]

    S. Chen, K. Saulnier, N. Atanasov, D. D. Lee, V . Kumar, G. J. Pappas, and M. Morari. Approximating explicit model predictive control using constrained neural networks. In Proc. of the 2018 American Control Conference , pages 1520–1527, 2018

    work page 2018

  11. [11]

    Teichrib and M

    D. Teichrib and M. Schulze Darup. Piecewise regression via mixed-integer programming for MPC. In Proceedings of the 6th Annual Learning for Dynamics & Control Con- ference, volume 242 of Proceedings of Machine Learning Research, pages 337–348. PMLR, 2024

    work page 2024

  12. [12]

    Xiang, H

    W . Xiang, H. D. Tran, J. A. Rosenfeld, and T. T. John- son. Reachable set estimation and safety verification for piecewise linear systems with neural network controllers. In Proc.of the 2018 American Control Conference , pages 1574–1579, 2018

    work page 2018

  13. [13]

    Abadi et al

    M. Abadi et al. TensorFlow: Large-scale machine learni ng on heterogeneous systems, 2015. Software available from tensorflow.org

    work page 2015

  14. [14]

    PyTorch: An Imperative Style, High-Performance Deep Learning Library

    A. Paszke et al. Pytorch: An imperative style, high- performance deep learning library. arXiv:1912.01703v1 [cs.LG], 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1912

  15. [15]

    Fazlyab, M

    M. Fazlyab, M. Morari, and G. J. Pappas. Safety ver- ification and robustness analysis of neural networks via quadratic constraints and semidefinite programming. IEEE Transactions on Automatic Control, 67(1):1–15, 2022

    work page 2022

  16. [16]

    Schwan, C

    R. Schwan, C. N. Jones, and D. Kuhn. Stability verifi- cation of neural network controllers using mixed-integer programming. IEEE Transactions on Automatic Control , 68(12):7514–7529, 2023

    work page 2023

  17. [17]

    Karg and S

    B. Karg and S. Lucia. Stability and feasibility of neu- ral network-based controllers via output range analysis. In Proc.of the 2020 Conference on Decision and Control , pages 4947–4954, 2020

    work page 2020

  18. [18]

    S. V . Rakovic, E. C. Kerrigan, K. I. Kouramas, and D. Q. Mayne. Invariant approximations of the minimal robust positively invariant set. IEEE Transactions on au- tomatic control, 50(3):406–410, 2005

    work page 2005

  19. [19]

    Goodfellow, D

    I. Goodfellow, D. Warde-Farley, M. Mirza, A. Courville , and Y . Bengio. Maxout networks. In Proc. of the 2013 In- ternational Conference on Machine Learning , volume 28, pages 1319–1327, 2013

    work page 2013

  20. [20]

    Markolf and O

    L. Markolf and O. Stursberg. Polytopic input constrain ts in learning-based optimal control using neural networks. In Proc. of the 2021 European Control Conference , pages 1018–1023, 2021

    work page 2021

  21. [21]

    Teichrib and M

    D. Teichrib and M. Schulze Darup. Error bounds for max- out neural network approximations of model predictive control. IF AC-PapersOnLine, 56(1):10113–10119, 2023

    work page 2023

  22. [22]

    Bemporad and M

    A. Bemporad and M. Morari. Control of systems in- tegrating logic, dynamics, and constraints. Automatica, 35(3):407–427, 1999

    work page 1999

  23. [23]

    The MOSEK optimization toolbox for MAT- LAB manual

    MOSEK ApS. The MOSEK optimization toolbox for MAT- LAB manual. V ersion 10.0., 2022

    work page 2022

  24. [24]

    Blanchini

    F. Blanchini. Ultimate boundedness control for uncert ain discrete-time systems via set-induced lyapunov functions . IEEE Transactions on Automatic Control , 39(2):428–433, 1994

    work page 1994

  25. [25]

    Lazar, D

    M. Lazar, D. Mu˜ noz de la Pe˜ na, W .P .M.H. Heemels, and T. Alamo. On input-to-state stability of min–max nonlin- ear model predictive control. Systems & Control Letters , 57(1):39–48, 2008

    work page 2008

  26. [26]

    L¨ uken and S

    L. L¨ uken and S. Lucia. Learning iterative solvers for a ccu- rate and fast nonlinear model predictive control via unsu- pervised training. In Proc. of the 2024 European Control Conference, pages 1843–1850, 2024

    work page 2024