Introduction
In modern control applications like autonomous driving or smart grids, systems must navigate environments where the "cost" of an action is often unknown and time-varying. While Online Convex Optimization (OCO) is a powerful tool for handling such uncertainty, it traditionally struggles with two critical aspects of physical systems: complex system dynamics and hard safety constraints. This paper introduces a novel framework that integrates OCO with Robust Model Predictive Control (MPC) to ensure that physical boundaries are never violated, even amidst disturbances and noisy sensor data.
The Challenge of Unknown Costs and Hard Constraints
Traditional control assumes a known cost function (e.g., minimizing fuel consumption). However, if an obstacle suddenly appears or energy prices fluctuate, the cost function changes unpredictably. Standard OCO algorithms typically allow for "soft" constraint violations, which is unacceptable in robotics or power systems where hitting a wall or overloading a circuit can be catastrophic. The authors address this by controlling a linear system facing unknown costs and external disturbances while guaranteeing strict adherence to safe operating limits.
Methodology: Robust OCO for Control
The proposed approach, Algorithm 1: Robust OCO for Control, merges MPC-style constraint tightening with the OCO loop. The process involves four key stages at each time interval:
- $\mu$-Step Prediction: Predicting the system's state several steps ahead based on historical inputs.
- Online Gradient Descent (OGD): Using the gradient of the previous cost function to estimate the current optimal steady state.
- Constraint Tightening: Shrinking the allowed state and input space to account for future uncertainty from disturbances and noise.
- Steering & Scaling: Calculating an input sequence to move toward the estimated optimal state. If a sequence risks violating a tightened constraint, a scaling factor ($\beta_t$) is applied to ensure safety.
Numerical Case Study: Autonomous Vehicle Tracking
The framework was evaluated through a simulation of an autonomous vehicle using nonlinear kinematics. The scenario required the car to maintain speed on a two-lane road, detect a slower vehicle ahead, and eventually overtake it. Random measurement noise was added to velocity and position sensors, and external disturbances were modeled to test the system's robustness.
Key Results and Insights
The authors provide a rigorous proof that the "dynamic regret"—the performance gap between their controller and a theoretical benchmark with perfect foresight—is bounded by the variation of the cost functions and the magnitude of disturbances. This links the learning performance of OCO with the stability requirements of control theory.
In simulations, the controller successfully managed complex maneuvers: cruising at 120 km/h, slowing to 70 km/h to maintain a 55m safety buffer, and then accelerating to 124 km/h to overtake. Crucially, the system respected all road boundaries and actuator limits despite the noisy sensor data. By utilizing Online Gradient Descent instead of solving large-scale optimization problems at every step, the controller remains computationally efficient for real-time use.
Limitations and Future Work
While the simulation demonstrated success with a nonlinear vehicle model, the formal theoretical proofs are currently limited to Linear Time-Invariant (LTI) systems. Future research aims to generalize these results to nonlinear or time-varying linear systems and explore more aggressive performance benchmarks, such as comparing against the optimal sequence of states in hindsight rather than just the optimal steady state.
Summary updated