Neurocontroller using dynamic state feedback for compensatory control

Csaba Szepesvári, Szabolcs Cimmer, András LörinczNeural Networks 1997 · DOI

TL;DR

This paper introduces a novel "Static and Dynamic State feedback" (SDS) control scheme that uses a neural network trained to approximate a plant's inverse dynamics. This approach enhances robustness against plant perturbations and unmodeled dynamics by employing a dual-purpose controller, leading to more stable and precise control, especially in scenarios with changing plant characteristics.

What Problem Does This Paper Address?

Traditional neurocontrol methods often rely on learning an inverse model of the plant for static state feedback. While effective, these methods are notoriously sensitive to even small errors in the inverse model or changes in the plant's dynamics, which can lead to instability. Existing solutions often involve complex adaptive control strategies or separate feedback controllers, which can introduce delays or credit assignment problems. This paper addresses the challenge of achieving robust and stable control in the presence of plant uncertainties and dynamic changes by proposing a unified approach that leverages the learned inverse dynamics for both feedforward and compensatory feedback control.

Methods

The core of the proposed method is the "Static and Dynamic State feedback" (SDS) controller. This controller utilizes two identical copies of a neural network trained to approximate the plant's inverse dynamics. One copy acts as the standard feedforward controller, generating the control signal based on the desired plant state. The second, identical copy acts as a feedback controller, generating a compensatory signal based on the actual plant state. This compensatory signal is then integrated over time and added to the feedforward signal.

The key innovation lies in how this compensatory signal is generated. It's derived from the difference between the desired plant dynamics and the actual plant dynamics, effectively correcting for errors and perturbations. The paper proves that under certain conditions (specifically, if the perturbation is "signproper"), this SDS scheme can ensure that the tracking error remains uniformly ultimately bounded, meaning it can be kept arbitrarily small by adjusting the feedback gain. The stability analysis is based on a modification of Liapunov's second method.

For learning, the paper favors direct inverse modeling, where the neural network learns the inverse dynamics from input-output data. This approach is compatible with the SDS scheme, allowing for simultaneous learning and feedback control.

Datasets

Not discussed in the paper. The paper focuses on theoretical development and simulations of a simplified sensorimotor loop and a robotic manipulator.

Key Insights and Ideas

  • Dual-Role Controller: Using the same learned inverse dynamics model for both feedforward and feedback control simplifies the system and addresses the credit assignment problem in learning.
  • Robustness through Dynamic Feedback: The dynamic state feedback component actively compensates for perturbations and unmodeled dynamics, significantly improving stability compared to purely static feedback.
  • "Signproper" Perturbations: The stability of the SDS controller is guaranteed if the perturbation is "signproper," a condition related to the angle between the effects of control on the perturbed and unperturbed plants. This is a crucial theoretical insight.
  • Error Compensation: The SDS controller can compensate for structural approximation errors in the learned inverse dynamics, making it more forgiving of imperfect learning.
  • Potential for Higher-Order Systems: The framework can be extended to higher-order plants, though careful design of the speed field is required.

Key Results

The paper presents theoretical proofs for the uniform ultimate boundedness of the tracking error under specific conditions. Simulations demonstrate the effectiveness of the SDS controller in compensating for:

  • Additive Perturbations: The compensated plant trajectory closely follows the desired path, deviating significantly from the uncompensated trajectory.
  • Multiplicative Perturbations: Similar to additive perturbations, the SDS controller significantly improves tracking accuracy.
  • Inhomogeneous Perturbations: Even with perturbations that vary across the state space (e.g., strongest at the center), the SDS controller shows improved performance, though overcompensation can occur due to integration time and strong non-linearities.

The simulations also suggest that the gain of the feedback component (A) should be chosen proportionally to the magnitude of the perturbation (e.g., the mass of an object held by a robot arm).

Limitations and Future Work

  • Signproperness Requirement: The theoretical guarantee of stability relies on the perturbation being "signproper." The paper acknowledges that detecting and handling non-signproper perturbations (which can lead to growing errors) is a significant open question.
  • Noise Sensitivity: While the paper discusses noise, it notes that noise entering the system just before the compensatory signal is integrated can be problematic, potentially making the system unstable and increasing the perturbation's amplitude with increasing gain.
  • Fast Actions: For very fast control actions, the feedback component might not have enough time to react. The paper suggests using feedback during the learning phase for such scenarios.
  • Higher-Order Plants: While the theory can be extended, designing appropriate speed fields for higher-order plants can be complex due to potential non-unique inverse dynamics.
  • Adaptive Gain: The paper suggests exploring adaptive or state-dependent gains to improve robustness to noise and varying plant dynamics.
  • Non-Stationary Perturbations: The current proofs assume stationary perturbations. While slow changes might be handled, fast, non-stationary perturbations can lead to instability. Further research is needed to address this.

Summary updated

BibTeX
@article{szepesvari1997neurocontroller,
  author = {Csaba Szepesvári and Szabolcs Cimmer and András Lörincz},
  title = {{Neurocontroller using dynamic state feedback for compensatory control}},
  journal = {Neural Networks},
  year = {1997},
  doi = {10.1016/s0893-6080(97)00043-9},
}
Figure