Abstract
Integrating black-box AI predictions into physical control systems offers significant performance potential but introduces risks of catastrophic failure when predictions are inaccurate. This paper presents a mathematical framework for Linear Quadratic Control (LQC) that balances performance with untrusted predictions (consistency) against worst-case safety guarantees (robustness). The authors propose a self-tuning controller that adaptively learns a trust parameter online, achieving near-optimal performance when predictions are accurate while maintaining a bounded competitive ratio regardless of prediction error.
Introduction
In modern engineering, there is a growing trend to incorporate high-precision neural network predictions into systems like power grids and autonomous robots. However, these "black-box" models often lack formal uncertainty quantification. In safety-critical environments, a single bad prediction can lead to system instability.
Classical control theory typically relies on Robust Control, which assumes the worst-case scenario to ensure safety but often sacrifices performance. Conversely, ML-augmented algorithms strive for Consistency—optimal performance under perfect predictions—but can fail spectacularly if those predictions are flawed. This research bridges the gap by developing a controller that uses untrusted predictions without risking an unbounded competitive ratio.
Methodology
The authors introduce two primary innovations to solve the tension between robustness and consistency in the LQC framework.
1. The $\lambda$-Confident Policy
Rather than choosing between fully trusting or completely ignoring AI predictions, the researchers propose a linear combination governed by a trust parameter $\lambda \in [0, 1]$.
- $\lambda = 1$: The controller fully trusts the AI predictions.
- $\lambda = 0$: The controller ignores predictions, reverting to a standard robust control policy.
The paper provides a theoretical bound on the Competitive Ratio (CR)—the ratio of the algorithm's actual cost to the offline optimal cost. This shows a predictable trade-off: higher $\lambda$ improves performance with good data but increases risk with bad data.
2. Self-Tuning Control
To handle scenarios where the accuracy of the AI is unknown, the authors developed an adaptive algorithm using a "Follow the Leader" (FTL) approach. The controller monitors past performance in real-time. If predictions were historically helpful, it increases $\lambda$; if they were misleading, it decreases trust. This self-tuning mechanism ensures the competitive ratio remains bounded even if prediction errors are massive, effectively learning the optimal balance of trust on the fly.
Experimental Evaluation
The framework was tested across three distinct scenarios to validate its theoretical guarantees:
- Robot Tracking: A 2D simulation where a robot followed complex trajectories. The self-tuning policy successfully navigated varying levels of binomial and Gaussian noise.
- Adaptive EV Charging: Using real-world data from Caltech’s ACN-Data, the model balanced battery health with fluctuating energy demands. The self-tuning policy performed nearly as well as the best possible "hindsight" trust parameter.
- Cart-Pole: To test practical transferability, the algorithm was applied to this classic nonlinear benchmark. Despite being designed for linear systems, the self-tuning trust parameter provided significant performance gains and safety through linearization.
Key Insights and Results
- The Lower Envelope Effect: Experiments demonstrated that the self-tuning policy effectively tracks the "lower envelope" of all possible fixed-$\lambda$ choices, finding the best trust level for the current error regime.
- Theoretical Robustness: The authors proved that the self-tuning policy achieves a competitive ratio that stays bounded regardless of how large the prediction error becomes. This prevents the cost from exploding during "black swan" prediction failures.
- Non-linear Error Dependency: The self-tuning mechanism creates a non-linear dependency on error, which is more resilient than standard linear error-correction methods.
Limitations and Future Work
While the results are promising, the authors note that regret bounds are not yet "tight" in regimes with extremely high and rapid variation in perturbations. Future work may involve adding regularizers (moving toward "Follow the Regularized Leader") to improve convergence in volatile settings. Additionally, while the Cart-Pole experiment showed success, extending these formal robustness and consistency guarantees to fully nonlinear control systems remains an open mathematical challenge.
Summary updated