Neural Network Based Approach for Tuning Kalman Filter

Abstract

Kalman filter (KF) parameter tuning has been dealt with in a limited fashion and usually was left to engineering intuition due to unavailable measurements of process noise and high dimensionality of the problem. In this paper we present a simple Neural Network (NN) based approach to KF tuning problem. Since the approach trades number of KF runs required for the optimal filter tuning for KF performance, the result of the such tuning is the set of tuning parameters that gives suboptimal performance. Advantages of this approach are: 1) simple practical framework for optimal filter performance tuning, 2) the framework is independent of the type of a filter and 3) low number of filter runs required to obtain quasi optimal parameter set. The main disadvantage is the suboptimal filter performance that can be easily improved by increasing the number of filter runs. Two NN architectures were investigated, generalized regression neural network (GRNN) and regular radial basis networks (RBNN). RBNN showed much better performance for a given non-linear test function with a clear maximum peak. Performance measures along with computational efficiency for these methods were compared. A step-by-step tuning procedure is presented.