RNN-UKF: Enhancing Hyperparameter Auto-Tuning in Unscented Kalman Filters Through Recurrent Neural Networks
Zhengyang Fan,Dan Shen,Yajie Bao,Khanh Pham,Erik Blasch,Genshe Chen
DOI: https://doi.org/10.23919/fusion59988.2024.10706523
2024-01-01
Abstract:The Unscented Kalman Filter (UKF) stands out as a versatile and dynamic algorithm, celebrated for its prowess in estimating the states of nonlinear dynamical systems within uncertain environments. However, the accuracy of UKF state estimations hinges significantly on the thoughtful selection of pivotal hyperparameters, $\alpha, \beta$, and $\kappa$, which are integral in shaping the distribution of sigma points around the current state estimate. Prevailing methods for tuning these parameters encompass heuristic approaches such as arbitrarily fix $\alpha$ at 0.001, $\kappa$ at 0, and $\beta$ at 2 for Gaussian noise, though the efficacy of such rules heavily hinges on the intricacies of the specific problem. Alternatively, the grid search technique seeks to optimize these hyperparameters, but it can become computationally burdensome, particularly when the search space is extensive and intricate. To navigate these hurdles, this paper introduces the RNN-UKF algorithm-a pioneering strategy that leverages recurrent neural networks (RNNs) to autonomously fine-tune UKF hyperparameters. The inherent adaptability of RNNs is harnessed to dynamically adjust the $\alpha, \beta$, and $\kappa$ parameters of the unscented transformation during each state estimation step, all aimed at minimizing the root mean squared error (RMSE). Demonstrated through numerical simulations, we provide compelling evidence that the RNN-UKF approach outperforms both heuristic rule of thumb and grid search techniques in terms of RMSE performance. Moreover, the RNN-UKF methodology showcases its superiority over the conventional extended Kalman filter (EKF) approach, particularly in scenarios characterized by substantial system noise.