Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

On the Generalization of Data-Assisted Control in port-Hamiltonian Systems (DAC-pH)

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Splitting at a virtual port lets RL learn only the dissipative half

desk verdict A genuinely interesting modular-control research proposal that is not yet a results paper: the key stability argument is assumed, the algebra has fixable errors, but the authors are transparent about the hypotheses. read the letter →

arxiv 2506.07079 v1 pith:HKJLPZQV submitted 2025-06-08 eess.SY cs.AIcs.LGcs.SY

classification eess.SYcs.AIcs.LGcs.SY MSC 93C1093C4093D0568T05
keywords Data-AssistedControlport-Hamiltoniansystemsreinforcementlearningpassivity-basedrobustadaptivevirtualportdecompositionbarrierfunctionssamplecomplexity
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 proposes a hypothetical hybrid control architecture for port-Hamiltonian (pH) systems, a broad energy-based modeling class used across mechanical, electrical, and networked physical systems. The dynamics are split at a virtual port $\Pi$ into a conservative, model-based part (the LHS) and a dissipative/input part (the RHS): the LHS evolves as $\dot{x}=J(x,\theta)\nabla H(x,\theta)+\Pi$, and the RHS realizes $\Pi=-R(x,\theta)\nabla H(x,\theta)+g(x,\theta)u$. A nonlinear robust adaptive controller commands $\Pi$ on the LHS, while a reinforcement learning agent learns to map the commanded port value to the actual actuator input on the RHS. The paper argues that this split confines the learning problem to residual dissipation and input uncertainty, preserves the pH energy structure, and enables modular safety and stability analysis. If correct, the framework would be a general template for structure-preserving learning-based control, and the authors support it with four hypotheses, preliminary analysis, and a pendulum simulation.

What carries the argument

The central object is the virtual port $\Pi$, placed at the interface between the conservative LHS and the dissipative/input RHS. It plays two roles at once: the RHS produces it through the physical map $\Pi=-R(x,\theta)\nabla H(x,\theta)+g(x,\theta)u$, and the LHS consumes it as an external input in $\dot{x}=J(x,\theta)\nabla H(x,\theta)+\Pi$. Commanding $\Pi$ with a robust adaptive controller and then learning a policy $u=\pi(\Pi_c,x)$ that realizes that command decouples energy shaping from uncertainty compensation. The structural assumptions—skew-symmetric $J$, positive-definite $R$, invertible $g$, differentiable $H$ with known gradient—are what make the RHS target static and affine in $u$ rather than a full temporal nonlinear policy.

What would settle it

Construct or find a pH system whose interconnection topology or Hamiltonian changes structurally during operation—for example, a circuit with a switched topology or a robot with unmodeled joint coupling—and check whether an LHS controller designed from fixed $J,H$ fails while an end-to-end learned controller succeeds; that would contradict the central decomposition claim. A softer quantitative test is to measure the sample complexity of the RHS policy against an end-to-end RL policy of matched capacity on a benchmark pH system; if DAC-pH does not require fewer samples for the same tracking error, Hypothesis 1 is contradicted.

Watch

Extended reading notes

Core claim

The paper's central claim is that a port-Hamiltonian system can be decomposed at a virtual port into a conservative LHS and a dissipative/input RHS, with the property that controlling the port variable controls the total energy dynamics. Because the interconnection matrix $J$ and Hamiltonian $H$ are assumed known up to slowly varying parameters, the LHS admits Lyapunov-based robust adaptive control and state/parameter estimation; because structural uncertainty is confined to $R$ and $g$, the RHS policy is a smaller, static, physically constrained target than an end-to-end policy. The paper therefore claims reduced hypothesis-class complexity and sample complexity relative to end-to-end RL, policy reusability across LHS controllers, and a path to closed-loop guarantees through a small-gain condition $\gamma_2\circ\gamma_1(s)<s$, approximated by $k_L k_R<1$. All of this is presented as a set of hypotheses with phenomenological, empirical, and preliminary theoretical support, not as proven theorems.

Load-bearing premise

The load-bearing premise is that the conservative half of the system—the interconnection topology $J$ and the Hamiltonian $H$—is known up to slowly varying parameters, so all structural uncertainty sits in the dissipation matrix $R$ and input map $g$; if that fails, the LHS controller cannot be synthesized and the framework's guarantees collapse.

Editorial extensions

If this is right

  • Reinforcement learning on the RHS must approximate only a residual map from the commanded port value to actuator input, so sample complexity and hypothesis-class size should shrink relative to end-to-end RL.
  • A learned RHS policy can be reused with different LHS controllers by changing the gain matrix $K_s$, because the two subsystems interact only through $\Pi$.
  • The LHS prior knowledge and the Hamiltonian support observers and parameter estimators under partial observability, improving the estimates the RHS policy relies on.
  • Safety can be layered: hard constraints through control barrier functions and a quadratic program on the applied input, soft constraints through reward shaping on proximity to boundaries.
  • Stability of the full loop can be approached by small-gain/ISS arguments, with the LHS gain $K_s$ as a tunable knob on the loop gain.

Reading between the lines

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

  • A concrete testable prediction, not run in the paper, is that on a matched benchmark the RHS policy under DAC-pH needs strictly fewer samples than an end-to-end RL policy with the same network capacity to reach comparable tracking error.
  • The same virtual-port split could apply to any control-affine system whose known part is conservative and whose uncertain part is dissipative/input, not only to pH systems; the paper only gestures at this generalization.
  • The boundary of the framework is the fixed-topology assumption on $J$ and $H$; if structural uncertainty reaches into the conservative part, the model-based LHS would itself need learning or adaptation, which the current hypotheses do not cover.
  • The pendulum experiments suggest that feeding the port discrepancy $\Pi-\Pi_c$ into the reward instead of the state error accelerates learning; turning that observation into a general reward-design principle would require the uniqueness argument to be made quantitative.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. This paper proposes a modular control framework, DAC-pH, for port-Hamiltonian systems. The pH dynamics (1) are split algebraically into a conservative part (LHS) driven by a virtual input Pi and a dissipative/input part (RHS) that realizes Pi from the physical control u. A model-based robust/adaptive controller is designed for the LHS, and reinforcement learning is proposed for the RHS. Four hypotheses are introduced concerning sample-complexity reduction, LHS stabilization, RHS/closed-loop stability and safety, and persistence-of-excitation-based learning. A pendulum simulation is presented as a proof of concept. The paper is explicitly framed as a hypothesis-generating contribution, with formal proofs deferred to future work.

Significance. The conceptual idea of separating known conservative dynamics from uncertain dissipative and actuation maps is appealing and could, if rigorously realized, yield interpretable, sample-efficient learning controllers for pH systems. The paper also provides a useful survey of structure-preserving learning methods and honestly labels its main statements as hypotheses. However, the paper does not currently provide a theorem or a validated design procedure; the preliminary stability analysis contains algebraic and logical errors, and the simulation is not a validation of the claimed safety and stability guarantees.

major comments (4)
  1. [Section 3, Hypothesis 2, Eq. (8)] With the control law (5), e = x - xd satisfies edot = -Ks e, so V1dot = -1/2 e^T(Ks^T P + P Ks)e, not e^T(Ks^T P + P Ks)e as written in Eq. (8). The factor 1/2 and the sign are missing. More importantly, the statement that Ks^T P + P Ks is negative definite if and only if Ks is Hurwitz is false for an arbitrary symmetric positive definite P; it holds only if P is a Lyapunov matrix for Ks, i.e., P solves the Lyapunov equation with some positive definite W. Since P in Eq. (6) is arbitrary, the asymptotic-stability conclusion is not established.
  2. [Section 3, Hypothesis 3, Preliminary Theoretical Justification] The stability proof assumes ||ePi|| <= beta(||ePi(0)||, t) on the grounds that policy improvement in RL is monotonic and the Bellman operator is a contraction. This is circular: convergence of value iteration or policy evaluation in expected-return space does not imply convergence of the realized port-matching error ePi = Pi - Pi_c, which depends on the learned policy, the uncertain maps R(.) and g(.), and the closed-loop trajectory. Moreover, the RHS is defined by the algebraic map Pi = -R grad H + g u, so no dynamics for ePi are specified; the term ePi^T edotPi in the derivative of V_DAC is therefore not well-defined without additional filter or state assumptions. The small-gain condition gamma2 o gamma1(s) < s is asserted, not derived.
  3. [Section 3, Hypothesis 2, Case 2, Eqs. (10)-(16)] The adaptive tracking argument relies on the 'regressor matching assumption' J(x,theta) grad H(x,theta) - J(x,theta_hat) grad H(x,theta_hat) approx -phi(x) theta_tilde, which is not justified for general pH systems and is exactly the kind of structural assumption the framework claims to avoid on the RHS. In addition, the V2 derivative calculation omits the port error ePi and the feedforward xd_dot terms from the error dynamics, and no condition on the time-varying gain Ks(t), such as a bound on its derivative, is given. Consequently, the conclusion V2dot = -e^T W e < 0 does not follow.
  4. [Section 3, opening paragraph, and Section 5] The paper states that the hypotheses 'will be systematically validated through ongoing theoretical and empirical investigations in subsequent research' and that 'future work will focus on providing rigorous theoretical foundations.' Since the central claims of the abstract and introduction, namely closed-loop stability and robustness guarantees and reduced learning complexity, are presented as hypotheses with phenomenological support rather than as proven statements or reproducible experimental results, the manuscript does not currently deliver the guarantees it advertises.
minor comments (6)
  1. [Abstract and Section 4.2] There are several typographical errors, including 'p$\mathcal{H}$' in the abstract, 'continues RL design' in Section 4.2, and 'state constratint' in Eq. (29). These should be corrected.
  2. [Section 3, Hypothesis 1, Eq. (4)] The sample-complexity bound in Eq. (4) is garbled; it should read m >= O((Pdim(H) + log(1/delta)) / epsilon^2) or similar, with the parentheses placed correctly.
  3. [Figure 8] The axis labels in Figure 8 appear corrupted (e.g., '&c' and '='), presumably representing Pi_c and tau. The labels should be fixed for readability.
  4. [Section 3, Hypothesis 4, Eq. (22)] The regressor phi(tau) in the persistence-of-excitation condition (22) is not defined precisely for the general pH setting; the authors should specify its construction from x, theta, and their estimates.
  5. [Section 3, Hypothesis 3] The claim that convergence x(t) -> xd(t) is equivalent to Pi(t) -> Pi_c(t) requires precise conditions such as identical initial conditions, absence of disturbances, and exact knowledge of J and H; as stated, the equivalence is too strong.
  6. [Section 4.3] The text says Figure 9 demonstrates a monotonic decrease in penalties, but the plotted average reward is increasing toward zero (less negative). The wording should be clarified.

Circularity Check

2 steps flagged · score 5.0 of 10

Stability guarantee in Hypothesis 3 is conditional on assuming the very eΠ-decay it aims to prove; sample-complexity reduction restates the decomposition.

  1. other [Section 3, Hypothesis 3, 'Preliminary Theoretical Justification']
    "Due to policy improvement in RL is monotonic decrease and Bellman operator is a contraction map, it is assumed that for some class KL function β(., .) we have ∥eΠ∥ ≤ β(∥eΠ(0), t∥), for all t≥0. Then according to converse Lyapunov Theorem, there exists class K function α(˙) such that eTΠ ˙eΠ ≤ −α(∥eΠ∥)."

    The preliminary stability proof needs the realized virtual-port mismatch eΠ = Π − Πc to decay, and it assumes exactly that decay ('it is assumed ... ∥eΠ∥ ≤ β(∥eΠ(0), t∥)') with Bellman contraction as support. Bellman contraction is a statement about value-function iteration in expectation; it does not by itself imply a KL bound on the closed-loop trajectory error eΠ, which depends on the learned policy, the uncertain R/g maps, and the state trajectory. Thus the small-gain condition γ2∘γ1(s)<s is not derived from RL properties but is an input assumption equivalent to the desired convergence. The paper labels this a hypothesis, yet the 'preliminary theoretical justification' uses the assumption as the load-bearing step of the stability conclusion.

  2. self definitional [Section 3, Hypothesis 1, 'Preliminary Theoretical Justification - Complexity of Learning']
    "Because the LHS handles known dynamics and state estimation, the RHS only needs to learn the residual or uncertain parts. This constrained structure inherently limits the size and complexity of the function class needed for the RHS."

    The claimed sample-complexity advantage of DAC-pH over end-to-end learning is obtained by defining HRHS as the residual/static map class and HE2E as the full policy class, and then asserting that Pdim(HRHS) is smaller. This is a restatement of the decomposition rather than a derived bound: no concrete HRHS, HE2E, or pseudo-dimension values are computed, and inequality (4) is only quoted as a generic PAC bound. The paper explicitly frames this as a hypothesis, but the 'reduction' is a definitional property of how the learning targets are chosen, not an independently established result.

full rationale

The paper is self-described as a hypothetical framework and repeatedly defers rigorous proofs to future work (e.g., 'Future work will focus on providing rigorous theoretical foundations for the proposed hypotheses, transforming them into formally proven theorems'). This lowers the severity of the circularity, because the hypotheses are not presented as fully derived theorems. Still, the central stability discussion in Hypothesis 3 assumes the very eΠ-decay property that the closed-loop stability claim needs, making the 'preliminary justification' circular in its load-bearing step. The sample-complexity claim in Hypothesis 1 is largely a definitional restatement of the decomposition, though it is also labeled a hypothesis. The decomposition (2) is an algebraic identity and is presented as a design choice rather than a derived structural property, so it is bookkeeping rather than circular. Self-citation of the authors' prior DAC work [1,2] is motivational, not load-bearing for the pH-specific stability argument. The pendulum simulation uses a known dissipative constant c and demonstrates feasibility, not a predictive validation. Overall, the paper's own caveats prevent a score in the 8-10 range, but the assumed eΠ-decay makes the claimed stability guarantee partially circular, yielding a score of 5.

Assumptions & free parameters 3 free parameters · 9 assumptions · 1 invented entities

The framework's central claims rest on the assumed pH structure, the invariant topology of J, the confinement of structural uncertainty to R and g, and several ad hoc assumptions in the stability and learning arguments (regressor matching, small-gain condition, PE). The only invented quantity is the virtual port Pi, which is a mathematical construct with no independent evidence.

free parameters (3)
  • LHS gain matrix Ks = k21=-0.5, k2=-1.5 (pendulum example)
    Chosen by hand to set eigenvalues; the framework treats Ks as a tunable parameter and as an RL action in some cases.
  • Integrator gain KI = [[-2,-1.5],[-1.5,-2]]
    Chosen to remove steady-state error in the pendulum example; not systematically tuned.
  • Reward weights wi = w1=1, w2=0.1, w3=0.1, w4=1
    Hand-selected normalization weights in Eq. (29); affect the learned policy trade-off.
assumptions (9)
  • standard math pH structure with skew-symmetric J and positive semidefinite R
    Eq. (1) and Section 2.1.
  • domain assumption Topology J and Hamiltonian structure of LHS are known and invariant
    Section 2.2: 'assuming the system's fundamental topology remains constant.'
  • domain assumption Structural uncertainty exists only in R and g, not J or H
    Section 3: 'we are assuming structural uncertainty on R and g present.'
  • standard math H is C^2 and J locally Lipschitz for uniqueness of solutions
    Hypothesis 3.
  • ad hoc to paper Regressor matching for adaptive law
    Hypothesis 2, Case 2: J(x,theta) grad H(x,theta) - J(x,theta_hat) grad H(x,theta_hat) approx -phi(x) theta_tilde.
  • domain assumption RL Bellman operator is a contraction and policy improvement monotone
    Hypothesis 3, used to assert ePi has KL decay.
  • ad hoc to paper Small-gain condition gamma2 o gamma1(s) < s and kL kR < 1
    Hypothesis 3, asserted for interconnection stability.
  • domain assumption Observability via topological coupling
    Hypothesis 2, Case 3: 'structural interconnection encoded in J(x) is assumed to enable the estimation of unmeasured states through topological coupling.'
  • domain assumption Persistency of Excitation with constants alpha1, alpha2, T
    Hypothesis 4, Eq. (22).
invented entities (1)
  • Virtual port variable Pi
    purpose: Interface between conservative LHS and dissipative/input RHS; serves as the command signal for LHS control and the target for RL policy.
    Defined as Pi = -R grad H + g u (Eq. 2); not directly measurable, must be estimated. It is a bookkeeping variable rather than a new physical entity, but the framework treats it as a control interface.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the Generalization of Data-Assisted Control in port-Hamiltonian Systems (DAC-pH)." pith.science (2026). https://pith.science/paper/HKJLPZQV

@misc{pith2026250607079,
  author       = {Pith},
  title        = {Pith review of: On the Generalization of Data-Assisted Control in port-Hamiltonian Systems (DAC-pH)},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HKJLPZQV}},
  note         = {Machine review of arXiv:2506.07079}
}
abstract

This paper introduces a hypothetical hybrid control framework for port-Hamiltonian (p$\mathcal{H}$) systems, employing a dynamic decomposition based on Data-Assisted Control (DAC). The system's evolution is split into two parts with fixed topology: Right-Hand Side (RHS)- an intrinsic Hamiltonian flow handling worst-case parametric uncertainties, and Left-Hand Side (LHS)- a dissipative/input flow addressing both structural and parametric uncertainties. A virtual port variable $\Pi$ serves as the interface between these two components. A nonlinear controller manages the intrinsic Hamiltonian flow, determining a desired port control value $\Pi_c$. Concurrently, Reinforcement Learning (RL) is applied to the dissipative/input flow to learn an agent for providing optimal policy in mapping $\Pi_c$ to the actual system input. This hybrid approach effectively manages RHS uncertainties while preserving the system's inherent structure. Key advantages include adjustable performance via LHS controller parameters, enhanced AI explainability and interpretability through the port variable $\Pi$, the ability to guarantee safety and state attainability with hard/soft constraints, reduced complexity in learning hypothesis classes compared to end-to-end solutions, and improved state/parameter estimation using LHS prior knowledge and system Hamiltonian to address partial observability. The paper details the p$\mathcal{H}$ formulation, derives the decomposition, and presents the modular controller architecture. Beyond design, crucial aspects of stability and robustness analysis and synthesis are investigated, paving the way for deeper theoretical investigations. An application example, a pendulum with nonlinear dynamics, is simulated to demonstrate the approach's empirical and phenomenological benefits for future research.

Figures

Figures reproduced from arXiv: 2506.07079 by the authors.

Figure 1
Figure 1. DAC-pH decomposition, environment, and uncertainty. d is disturbance input and n is noise. Π acts as a generalized momentum/flow that the LHS subsystem uses as an external input and that the RHS subsystem generates based on u. In an ideal decomposition, we would design a controller to specify Π for the LHS subsystem, and then separately solve for the actual input u that produces that Π. This decomposition is motivat… view at source ↗
Figure 2
Figure 2. Examples of physical system decomposition boundaries and [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Trajectory of state, input, and port value in DAC-p [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Illustration of soft and hard constraints in DAC-p [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]
Figure 5
Figure 5. Figure 5: DAC-pH system architecture - assuming Π = Πc, the LHS and RHS are virtually closed-loop LHS Virtual Closed - Loop RHS Virtual Closed-Loop Π Π𝑐 𝑒ƸΠ 𝑒𝑥 𝑥 [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: LHS and RHS virtual closed-loop systems in interconnection [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: SAC networks’ design: Each fully connected input layer ( [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: Pendulum system under DAC-pH control: 10 simulations for random initial condition of x1 ∈ [−π/4, π/4], x2 = 0 Average Reward with 95% Confidence Interval 0 5 10 15 20 25 30 35 40 45 Episode -4 -3.5 -3 -2.5 -2 -1.5 -1 -0.5 0 0.5 Av era g e R e w ard #10 5 95% CI Average…
Figure 9
Figure 9. Figure 9: Moving average reward and its corresponding confidence level over training episodes, using a window size of [PITH_FULL_IMAGE:figures/full_fig_p021_9.png]
Figure 10
Figure 10. Figure 10: Effect of state feedback gain k21 and k2 on overall performance To eliminate the residual steady-state error, integrators are added to the control law as follows: z˙1 = x1, z˙2 = x2, Πc = −J∇H(x) + Ksx + KI z, (30) where z = [z1, z2] ⊤ represents the integrator states…
Figure 11
Figure 11. Figure 11: Effect of LHS integrator random sampling and sensitivity analysis—are necessary to establish principled guidelines for hyperparameter tuning. • For large-scale systems, leveraging the LHS model to improve reward estimation and predict system trajectories can enhance b…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Learning-Based Data-Assisted Port-Hamiltonian Control for Free-Floating Space Manipulators

    eess.SY 2025-09 conditional novelty 4.0 of 10

    A controller that learns only the uncertain input side of a port-Hamiltonian space robot model keeps a simulated free-floating manipulator stable while tracking orientation.

Reference graph

Works this paper leans on

24 extracted references · 21 canonical work pages · cited by 1 Pith paper

  1. [1]

    Data-assisted control: A framework development by exploiting nasa generic transport platform.International Journal of Robust and Nonlinear Control, 34(3):1898–1920, 2024

    Mostafa Eslami and Afshin Banazadeh. Data-assisted control: A framework development by exploiting nasa generic transport platform.International Journal of Robust and Nonlinear Control, 34(3):1898–1920, 2024

  2. [2]

    Sequential data-assisted control in flight.arXiv, 2023

    Mostafa Eslami and Afshin Banazadeh. Sequential data-assisted control in flight.arXiv, 2023. 23 On the Generalization of Data-Assisted Control in port-Hamiltonian Systems (DAC-pH)

  3. [3]

    Optimal control of port-hamiltonian systems: A continuous-time learning approach.Automatica, 130:109725, 2021

    Lukas Kölsch, Pol Jané Soneira, Felix Strehle, and Sören Hohmann. Optimal control of port-hamiltonian systems: A continuous-time learning approach.Automatica, 130:109725, 2021

  4. [4]

    Port-hamiltonian systems with energy and power ports

    Kaja Krhaˇc, Bernhard Maschke, and Arjan van der Schaft. Port-hamiltonian systems with energy and power ports. IFAC-PapersOnLine, 58(6):280–285, 2024. 8th IFAC Workshop on Lagrangian and Hamiltonian Methods for Nonlinear Control LHMNC 2024

  5. [5]

    Model-free lqr design by q-function learning.Automatica, 137:110060, 2022

    Milad Farjadnasab and Maryam Babazadeh. Model-free lqr design by q-function learning.Automatica, 137:110060, 2022

  6. [6]

    Nageshrao, and Gabriel A

    Olivier Sprangers, Robert Babuška, Subramanya P. Nageshrao, and Gabriel A. D. Lopes. Reinforcement learning for port-hamiltonian systems.IEEE Transactions on Cybernetics, 45(5):1017–1027, 2015

  7. [7]

    Nageshrao, G.A.D

    S.P. Nageshrao, G.A.D. Lopes, D. Jeltsema, and R. Babuška. Passivity-based reinforcement learning control of a 2-dof manipulator arm.Mechatronics, 24(8):1001–1007, 2014

  8. [8]

    Adaptive h∞ control of uncertain port-controlled hamiltonian systems subject to actuator saturation

    Airong Wei, Yuzhen Wang, and Xiaoming Hu. Adaptive h∞ control of uncertain port-controlled hamiltonian systems subject to actuator saturation. InProceedings of the 30th Chinese Control Conference, pages 391–396. IEEE, 2011

Show all 24 references
  1. [9]

    Desai, Marios Mattheakis, David Sondak, Pavlos Protopapas, and Stephen J

    Shaan A. Desai, Marios Mattheakis, David Sondak, Pavlos Protopapas, and Stephen J. Roberts. Port-hamiltonian neural networks for learning explicit time-dependent dynamical systems.Phys. Rev. E, 104:034312, Sep 2021

  2. [10]

    Integrating port-hamiltonian systems with neural networks: From deterministic to stochastic frameworks.arXiv preprint arXiv:2403.16737, 2024

    Luca Di Persio, Matthias Ehrhardt, and Sofia Rizzotto. Integrating port-hamiltonian systems with neural networks: From deterministic to stochastic frameworks.arXiv preprint arXiv:2403.16737, 2024

  3. [11]

    Port-hamiltonian neural ode networks on lie groups for robot dynamics learning and control.IEEE Transactions on Robotics, 2024

    Thai Duong, Abdullah Altawaitan, Jason Stanley, and Nikolay Atanasov. Port-hamiltonian neural ode networks on lie groups for robot dynamics learning and control.IEEE Transactions on Robotics, 2024

  4. [12]

    Jiahao, and George J

    Thomas Beckers, Tom Z. Jiahao, and George J. Pappas. Learning switching port-hamiltonian systems with uncertainty quantification.IFAC-PapersOnLine, 56(2):525–532, 2023. 22nd IFAC World Congress

  5. [13]

    Springer Science & Business Media, 2009

    Vincent Duindam, Alessandro Macchelli, Stefano Stramigioli, and Herman Bruyninckx.Modeling and control of complex physical systems: the port-Hamiltonian approach. Springer Science & Business Media, 2009

  6. [14]

    Port-hamiltonian flight control of a fixed-wing aircraft.IEEE Transactions on Control Systems Technology, 30(1):408–415, 2021

    Jean-Michel Fahmi and Craig A Woolsey. Port-hamiltonian flight control of a fixed-wing aircraft.IEEE Transactions on Control Systems Technology, 30(1):408–415, 2021

  7. [15]

    Expressiveness and structure preservation in learning port-hamiltonian systems

    Juan-Pablo Ortega and Daiying Yin. Expressiveness and structure preservation in learning port-hamiltonian systems. In Frank Nielsen and Frédéric Barbaresco, editors,Geometric Science of Information, pages 313–322, Cham, 2023. Springer Nature Switzerland

  8. [16]

    Port-hamiltonian dynamic mode decomposition

    Riccardo Morandin, Jonas Nicodemus, and Benjamin Unger. Port-hamiltonian dynamic mode decomposition. SIAM Journal on Scientific Computing, 45(4):A1690–A1710, 2023

  9. [17]

    Springer International Publishing, Cham, 2017

    Arjan van der Schaft.Port-Hamiltonian Systems, pages 113–171. Springer International Publishing, Cham, 2017

  10. [18]

    Springer, 2011

    Wolfgang Borutzky.Bond graph modelling of engineering systems, volume 103. Springer, 2011

  11. [19]

    Bond-graph modeling.IEEE Control Systems Magazine, 27(2):24–45, 2007

    Peter J Gawthrop and Geraint P Bevan. Bond-graph modeling.IEEE Control Systems Magazine, 27(2):24–45, 2007

  12. [20]

    Concept-oriented modeling of dynamic behavior.Bond Graph Modelling of Engineering Systems: Theory, Applications and Software Support, pages 3–52, 2011

    Peter C Breedveld. Concept-oriented modeling of dynamic behavior.Bond Graph Modelling of Engineering Systems: Theory, Applications and Software Support, pages 3–52, 2011

  13. [21]

    MIT press Cambridge, 1998

    Richard S Sutton, Andrew G Barto, et al.Reinforcement learning: An introduction, volume 1. MIT press Cambridge, 1998

  14. [22]

    Cam- bridge university press, 2014

    Shai Shalev-Shwartz and Shai Ben-David.Understanding machine learning: From theory to algorithms. Cam- bridge university press, 2014

  15. [23]

    cambridge university press, 2009

    Martin Anthony and Peter L Bartlett.Neural network learning: Theoretical foundations. cambridge university press, 2009

  16. [24]

    Spectrally-normalized margin bounds for neural networks

    Peter L Bartlett, Dylan J Foster, and Matus J Telgarsky. Spectrally-normalized margin bounds for neural networks. Advances in neural information processing systems, 30, 2017. 24

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.