The Autoencoder-Kalman Filter: Theory and Practice
Matthew L. Weiss, Randy Clinton Paffenroth, Joshua R. Uzarski · 2019
Given a noisy signal, it is often of interest to estimate its noise-free state. One of the more common state estimation techniques is the Kalman Filter, which is optimal under certain conditions, one of which is that the measurements are a Gaussian random process with known covariance R. However, in practical applications the covariance may not be known or the noise process may not be Gaussian. Accordingly, here we employ a method called the Autoencoder-Kalman Filter (AEKF) to learn a mapping from noisy measurements to inputs for the Kalman Filter. Training the AEKF uses a technique called domain randomization and the AEKF has been shown to outperform the Kalman Filter and a LSTM neural network on a variety of noise types. However, up to this point the training has been limited to specific classes of functions. In this research, utilizing domain randomization, we propose to train an AEKF on random polynomial functions, which is able to accurately filter a larger class of smooth curves with multiple noise types.