This paper, presented at the 2021 IEEE Conference on Decision and Control (CDC), tackles a "reverse-engineering" problem in robust control: if you know where you want your system's outputs to land, how do you calculate the set of allowable input disturbances that will get you there?
TL;DR
The authors propose a novel optimization framework to compute input disturbance sets (polytopes) that result in a desired set of reachable outputs. By framing this as a bilevel optimization problem involving robust invariant sets, they provide a way to size actuators or define safe reference regions for constrained linear systems.
What Problem Does This Paper Address?
In robust control, researchers typically start with a known "disturbance set"—the bounds on the noise or uncertain inputs affecting a system—and then calculate the resulting "reachable set" of states or outputs. This helps ensure the system stays safe.
However, many engineering tasks require the opposite. For example:
- Actuator Design: How much range (input authority) does a motor need to ensure the robot can reach every point in a specific workspace?
- Supervisory Control: What is the largest set of commands a human operator can send without the system eventually hitting a physical constraint?
- System Identification: What is the smallest noise set that explains the experimental data we’ve observed?
The challenge is that the relationship between input sets and output sets is governed by Robust Positive Invariant (RPI) sets—regions where, if the system starts inside, it is guaranteed to stay inside despite disturbances. Computing these sets is mathematically complex, and optimizing over them to find an unknown disturbance set is even harder.
Methods
The authors represent the disturbance set as a polytope (a higher-dimensional polygon) with fixed "faces" but variable "offsets." Their approach involves several key steps:
- Bilevel Optimization: They set up a two-tier problem. The "upper" level tries to minimize the distance (Hausdorff distance) between the achieved output set and the target set. The "lower" level ensures that the state set used in the calculation is the tightest possible RPI set for a given disturbance.
- Minimal RPI Approximation: Since the true minimal RPI set is often impossible to compute exactly, they use a parameterized polytopic version that can be integrated into an optimization loop.
- Penalty Function Method: Solving bilevel problems is computationally expensive. The authors use a penalty function to collapse the two levels into a single-level optimization problem, which can then be solved using standard linear programming tools.
- Inner and Outer Approximations: The framework is flexible. It can find the largest disturbance set that stays within a safe output zone (inner) or the smallest disturbance set required to cover a target output zone (outer).
Datasets
The paper does not use external real-world datasets. Instead, it validates the method using two benchmark numerical simulations:
- An LQI-tracking system: A 4th-order system representing a controller tracking a reference signal while respecting input saturation limits.
- A Double-Integrator model: A standard 2nd-order system used to demonstrate output reachability.
Key Insights and Ideas
- The "Aha!" Moment on Uniqueness: A significant theoretical contribution (Theorem 1) proves that the fixed-point iteration used to find the RPI set is unique even under relaxed conditions. This ensures the optimization algorithm doesn't get "stuck" or produce ambiguous results.
- Distance as a Metric: By using the Hausdorff distance, the authors provide a mathematically rigorous way to say one set "looks like" another, allowing the optimizer to "shape" the disturbance set to match the target output.
- Practicality via Duality: By leveraging the duality of linear programs, they transform complex set-inclusion constraints into simple linear inequalities that a computer can solve quickly.
Key Results
- Safe Reference Sets: In the first example, the method successfully computed the maximum allowable range for a reference signal (e.g., a joystick input) such that the system's internal states would never violate physical constraints, even without the supervisor knowing the exact current state of the system.
- Reachability Coverage: In the second example, the algorithm found the smallest input set required to ensure that a double-integrator system could reach every point in a target output box. The results showed that the computed set tightly matched the target, demonstrating high accuracy.
Limitations and Future Work
- Fixed Hyperplanes: Currently, the user must choose the "shape" (the normal vectors of the polytope faces) of the disturbance set beforehand. The optimizer only adjusts the size/offset. Future work could involve optimizing the shape itself.
- Computational Scaling: While the penalty method is efficient, bilevel optimization can still become computationally heavy for very high-dimensional systems (e.g., hundreds of states).
- Alternative Solvers: The authors suggest exploring "value function" approaches or Newton-type procedures to handle the nonlinearities that arise when the shape of the polytope is also an optimization variable.
Summary updated