Finite Approximations of Switched Homogeneous Systems for Controller Synthesis

Danielle C. Tarraf, Alexandre Megretski, Munther A. DahlehIEEE Trans. Automatic Control 2011 · DOI

Abstract

Hybrid systems combining continuous dynamics and discrete switching are challenging to control, particularly when only coarse sensors are available. This paper presents a framework for stabilizing switched homogeneous systems by approximating them as deterministic finite state machines (DFMs). By leveraging coordinate transformations and small-gain theory, the authors provide a "correct-by-design" synthesis pipeline that guarantees stability despite approximation errors.

The Challenge of Coarse Control

Hybrid systems—which combine continuous "analog" dynamics with discrete "switching" logic—are notoriously difficult to control, especially when they are inherently unstable. Traditional control methods often rely on high-precision sensors, but in many real-world applications (like low-cost robotics or power-constrained satellites), we only have access to "coarse" data, such as a binary sensor that only tells us if a value is positive or negative.

Feedback control loop with binary input $u(t)$ and binary output $y(t)$.

Existing techniques for simplifying these systems (abstractions) often struggle with instability or produce models that are too complex for practical controller design. The authors aim to bridge this gap by creating a simplified, finite-state model that is computationally tractable while still providing a mathematical guarantee that the resulting controller will stabilize the original, complex system.

Constructive Methods for Abstraction

The authors propose a three-step constructive approach to bridge the gap between continuous dynamics and discrete control:

  1. Coordinate Transformation: They focus on "homogeneous" systems. By converting the system to polar coordinates, they exploit a key property: the stability of the system depends on the angle of the state, not its distance from the origin. This allows them to project the 2D problem onto a 1D unit circle.
  2. Finite State Machine (DFM) Construction: The unit circle is partitioned into a finite number of intervals. These intervals, and unions of adjacent intervals, become the states of a Deterministic Finite State Machine. This DFM acts as an "observer," tracking the plant's state based on binary feedback.
  3. Robust Synthesis via Small Gain Theory: Because the DFM is only an approximation, there is an "error" between the model and reality. The authors use a "small gain" argument to calculate a formal upper bound on this error using linear programming.
Approximation of the plant $P$ using a finite model $M$ and uncertainty block $\Delta$.

They then use dynamic programming (Value Iteration) to find a switching law that is robust enough to handle this error, ensuring the original plant converges to the origin.

Key Insights and Results

The "aha moment" of this research is realizing that for homogeneous systems, one only needs to track the "direction" of the state on the unit circle, drastically simplifying the approximation. Furthermore, by feeding the plant's output back into the finite state model, the authors prevent approximation errors from growing infinitely—a common pitfall for unstable systems.

The method was tested on a "harmonic oscillator" example, switching between a passive mode and an "aggressive" unstable mode using only the sign of the position as feedback. Key findings include:

  • Scalability: Increasing the number of intervals in the circle partition decreased the provable error bound and improved the guaranteed convergence rate.
  • Efficiency: A DFM with only 39 states was sufficient to stabilize a specific unstable configuration.
  • Robustness: Simulations showed that the actual convergence rate was often significantly better than the theoretical guaranteed rate, suggesting the method is robust even with conservative bounds.

Limitations and Future Work

While effective, the current method for calculating error bounds is "worst-case," which can lead to overly conservative controllers. Additionally, the "curse of dimensionality" remains a challenge; while 2D systems are well-handled, the state space grows exponentially for higher-order systems. Future research will focus on finding tighter bounds and extending these "gain condition" abstractions to a broader class of non-linear hybrid systems.

Summary updated

BibTeX
@article{tarraf2011finite,
  author = {Danielle C. Tarraf and Alexandre Megretski and Munther A. Dahleh},
  title = {{Finite Approximations of Switched Homogeneous Systems for Controller Synthesis}},
  journal = {IEEE Trans. Automatic Control},
  year = {2011},
  doi = {10.1109/tac.2011.2108350},
}
Figure