The Rationale for Second Level Adaptation

Kumpati S. Narendra, Yu Wang, Wei ChenarXiv 2015 · OpenReview

TL;DR

This paper delves into "second level adaptation," a novel adaptive control strategy that leverages multiple plant models to estimate unknown parameters more rapidly and robustly than conventional methods. The authors provide a theoretical and experimental examination of its advantages, particularly for systems with large uncertainties or time-varying parameters.

What Problem Does This Paper Address?

Adaptive control aims to control systems with unknown or changing parameters. Traditional methods, often called "first level adaptation," can struggle with large uncertainties or rapidly varying parameters, leading to slow convergence or instability. This paper addresses the limitations of first-level adaptation by exploring a "second level adaptation" approach that uses multiple models to create a more efficient and stable estimation of the plant's unknown parameters. The core challenge is to develop a method that can quickly and reliably identify these parameters to ensure effective control, even under difficult conditions.

Methods

The paper introduces and analyzes "second level adaptation," which builds upon the idea of using multiple models to estimate the unknown plant parameter vector, denoted as $\theta_p$. The key innovation lies in using $N \geq m+1$ models (where $m$ is the dimension of $\theta_p$) to estimate an alternative parameterization, $\alpha$, which is then used for control.

  • Multiple Models: The approach utilizes $m+1$ models, either fixed or adaptive, to estimate $\theta_p$. These models are initialized with different parameter values that span the convex hull of the possible $\theta_p$ values.
  • Alternative Parameterization ($\alpha$): Instead of directly estimating $\theta_p$, the method estimates a vector $\alpha$. This $\alpha$ is derived from the relationships between the parameter estimates from the multiple models and the true plant parameters. The paper shows that $\alpha$ can be estimated using linear equations, which is a significant advantage over the nonlinear equations often involved in direct $\theta_p$ estimation.
  • Estimation of $\alpha$: The paper details two methods for estimating $\alpha$:
  • Algebraic Estimation: Using linear equations derived from the model outputs.
  • Differential Equation Estimation: A more robust approach that uses a differential equation to update $\alpha$, providing better stability guarantees.
  • Control: Once $\alpha$ is estimated, it's used to determine the feedback control law, effectively controlling the plant through this alternative parameterization.

The paper contrasts this with "first level adaptation," which typically involves a single adaptive model and often relies on nonlinear equations for parameter estimation.

Datasets

Not discussed in the paper. The paper relies on simulated experiments to evaluate the proposed method.

Key Insights and Ideas

  • Linearity is Key: The core advantage of second level adaptation stems from its ability to estimate an alternative parameterization ($\alpha$) using linear equations, even when the original plant parameters ($\theta_p$) might be difficult to estimate non-linearly. This linear estimation of $\alpha$ significantly simplifies stability analysis and improves convergence speed.
  • Many Models, Faster Adaptation: Using $m+1$ models, where $m$ is the parameter dimension, allows for a much faster convergence of the parameter estimates compared to single-model approaches, especially when the parameter uncertainty is large.
  • Robustness to Time-Varying Parameters: The method shows improved performance and robustness when plant parameters change over time, outperforming conventional adaptive control in such scenarios.
  • Designer's Control: The structure of the estimation matrix $E(t)$ in the second level adaptation depends on the chosen identification models, not the unknown plant. This gives the designer more control over the adaptation process and its convergence rate.
  • Fixed vs. Adaptive Models: While adaptive models can be used, fixed models can also be effective, particularly for time-varying plants, as they don't need to track the parameter variations themselves. The choice can depend on the specific application.

Key Results

  • Faster Convergence: Simulation studies demonstrate that second level adaptation converges orders of magnitude faster than first level adaptation, with convergence times of 16 seconds versus barely commenced adaptation for first level methods in one example.
  • Reduced Tracking Error: The tracking error for second level adaptation converges much faster (e.g., in 3 seconds) compared to first level methods.
  • Improved Performance with Large Uncertainty: Simulations show significantly better performance for second level adaptation when dealing with large parameter uncertainties.
  • Robustness: The method is found to be more robust than first level adaptation in scenarios with time-varying parameters and potentially with noise, although the effect of unmodeled dynamics is still under investigation.
  • Theoretical Stability: The paper provides a proof of stability for the second level adaptation system, even when control is based on the estimated parameter $\hat{\alpha}$.

Limitations and Future Work

  • Plant Representation: The analysis is primarily conducted for plants in companion form with accessible state variables. Extensions to more general representations and adaptation using only input-output data are areas for future investigation.
  • Unmodeled Dynamics: The robustness of second level adaptation in the presence of unmodeled dynamics is an area currently being investigated.
  • Combining Models: Further research is needed to determine how fixed and adaptive models can be optimally combined for maximum effect in second level adaptation.
  • Model Location: While the paper notes that the location of models provides control over adaptation speed, further work is suggested to determine optimal model placement for maximal marginal benefit.

Summary updated

Abstract

Recently, a new approach to the adaptive control of linear time-invariant plants with unknown parameters (referred to as second level adaptation), was introduced by Han and Narendra in [1]. Based on $N (\geq m+1)$ fixed or adaptive models of the plant, where $m$ is the dimension of the unknown parameter vector, an unknown parameter vector $\alpha\in R^{N}$ is estimated in the new approach, and in turn, is used to control the overall system. Simulation studies were presented in [1] to demonstrate that the new method is significantly better than those that are currently in use. In this paper, we undertake a more detailed examination of the theoretical and practical advantages claimed for the new method. In particular, the need for many models, the proof of stability, and the improvement in performance and robustness are explored in depth both theoretically and experimentally.
BibTeX
@misc{narendra2015rationale,
  author = {Kumpati S. Narendra and Yu Wang and Wei Chen},
  title = {{The Rationale for Second Level Adaptation}},
  year = {2015},
  eprint = {1510.04989},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1510.04989},
}

Loading related content...

Figure