When are selector control strategies optimal for constrained monotone systems?

Hamed Taghavian, Ross Drummond, Mikael JohanssonarXiv 2023 · DOI · arXiv

TL;DR

This paper provides a rigorous mathematical justification for why simple "selector control" strategies—common in industrial PID setups—are actually optimal for a broad class of monotone systems. By proving that optimal solutions for these systems are "bang-ride" (always hugging the constraint boundaries), the authors show that complex optimization can often be replaced by a simple bank of controllers and a min/max selector.

What Problem Does This Paper Address?

In the world of control theory, there is a persistent gap between high-level "optimal control" (like Model Predictive Control) and what actually happens on the factory floor. While researchers often favor complex, memory-intensive optimization algorithms, industry overwhelmingly relies on PIDs and simple logic gates called selectors. A selector simply looks at several controller outputs and picks the smallest (or largest) one to send to the plant.

Despite their ubiquity in power grids, chemical plants, and engines, selectors are often implemented in an ad-hoc manner. We haven't had a clear theoretical understanding of when these simple strategies are truly optimal. This paper tackles that gap by identifying the specific properties of a system—specifically "monotone" systems with monotone constraints—that make selector control the mathematically best way to operate.

Methods

The authors use the framework of monotone dynamical systems. These are systems where, roughly speaking, if you start with a "larger" state or provide a "larger" input, the future states and outputs will also be "larger."

A typical selector control architecture where multiple PID controllers compete via a 'min' block to determine the plant input.

The key methodological contributions include:

  1. Bang-Ride Proofs: The authors prove that for monotone systems with monotone costs and constraints, the optimal control sequence is "bang-ride." This means the controller should always be pushing the system until it hits at least one constraint (either the maximum allowed input or a safety limit on an output).
  2. Impulse Response Analysis: For linear systems, they show that if the system's impulse response is decreasing over time, the bang-ride property holds.
  3. Feedback Synthesis: They demonstrate that this optimal "bang-ride" behavior can be mathematically represented as a nonlinear state-feedback law: $u_t = \min\{K_0(x_t), \dots, K_p(x_t)\}$. This structure is identical to a min-selector.
  4. Heuristic Transition: They bridge the gap to practice by showing how the theoretical state-feedback laws ($K_i$) can be replaced with standard PID controllers in a selector structure, as shown in the diagram above.

Datasets and Validation

The paper validates the theory using high-fidelity mathematical simulations of Lithium-ion battery models:

  • Single-Particle Models (SPM): A physics-based model describing lithium diffusion.
  • Equivalent-Circuit Models (ECM): An empirical model using resistors and capacitors to mimic battery behavior.

The authors use parameters for various battery chemistries (Lithium Cobalt Oxide, etc.) to demonstrate that real-world battery charging problems naturally fit their "monotone" criteria.

Key Insights and Ideas

  • The "Bang-Ride" Phenomenon: In many optimal control problems, you might expect the controller to gracefully navigate the state space. Here, the authors show that for monotone systems, the "best" path is almost always "riding" the edge of what is allowed.
  • Justifying the PID Selector: The most significant "aha moment" is the structural link. The paper proves that the optimal solution must look like a min-selector. This justifies why engineers have been using selectors for decades—they were intuitively gravitating toward the optimal structure for these types of systems.
  • Relative Degree Handling: The authors provide a clever way to handle systems where the input doesn't immediately affect the output (non-zero relative degree) by redefining the output functions as a forward time-shift, allowing the theory to apply to a much wider range of hardware.

Key Results

  • Theoretical Optimality: The paper successfully identifies two major families of problems (defined in Theorem 1 and Theorem 2) where selector control is guaranteed to be optimal.
  • Battery Fast-Charging: In simulations of health-aware battery charging (maximizing charge while staying under voltage and temperature limits), the heuristic PID selector performed almost identically to the theoretically optimal bang-ride controller.
  • Constraint Satisfaction: The authors show that using specific PID tuning rules (focused on first impulse-response samples) can significantly reduce constraint violations during the "switch" between different control loops (e.g., switching from constant current to constant voltage).

Limitations and Future Work

  • Model-Free Tuning: While the paper justifies the structure of selector control, tuning the individual PIDs in a model-free way that guarantees safety (no overshoots) remains a challenge.
  • Stability Guarantees: The paper focuses on optimality but notes that a general proof for the stability and feasibility of PID-based selectors (especially during the switching transients) is still an open question.
  • State Access: The theoretical optimal law requires full state measurements, which aren't always available. While PIDs bypass this, the "optimality gap" created by using output-feedback instead of state-feedback deserves further study.

Summary updated

BibTeX
@misc{taghavian2023when,
  author = {Hamed Taghavian and Ross Drummond and Mikael Johansson},
  title = {{When are selector control strategies optimal for constrained monotone systems?}},
  year = {2023},
  eprint = {2312.08829},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2312.08829},
}
Figure