We propose direct estimations of filter distributions in Kalman filter from data using density ratio estimation and KDE. Moreover, we apply the direct estimations to state estimations and propose new state estimations without information on measurement equations and noise.
