Influence Functions for Data Attribution in Linear System Identification and LQR Control

Abstract: Understanding the influence of individual training data points is crucial for developing reliable machine learning-based control systems. However, conventional methods like leave-one-out retraining are computationally infeasible for large datasets. This paper introduces a framework using influence functions to efficiently approximate the impact of removing specific training trajectories on both learned system dynamics and downstream control performance. We formulate two influence functions(IF): IF1, which estimates the effect on the predictive accuracy of a learned linear dynamics model, and IF2, which quantifies the subsequent impact on the cost of a Linear Quadratic Regulator (LQR) controller designed using these learned dynamics. These involve tracing sensitivities through the Discrete Algebraic Riccati Equation (DARE) solution. We empirically validate our approach on simulated linear systems analogous to robotic manipulators. Results show strong positive correlations between influence predictions and ground truth changes obtained via retraining. Our framework provides a computationally tractable method for data attribution.