Initial Conditions for Kalman Filtering: Prior Knowledge Specification

摘要

The paper deals with a selection of the initial state for Kalman filtering. The prior knowledge about it can be highly uncertain. In practice the initial state mean and covariance are often chosen arbitrarily. The present paper considers the problem from the position of knowledge elicitation and proposes a methodology to extract the prior knowledge from available information by the respective processing in order to choose the adequate initial conditions. The suggested methodology is based on utilization of the conjugate prior distribution for models, belonging to the exponential family.