Efficient Filter for Abruptly Changing Systems

摘要

For a linear discrete time system with coefficients that are governed by an N-state Markov chain, e.g., an aircraft tracked by surveillance radar, a filtering algorithm is defined. It consists of a bank of N interacting Kalman-like filters which communicate with a filter for the Markov chain. When the transition probabilities of the Markov chain are zero, the interacting multiple model (IMM) algorithm reduces to the MM algorithm. Qualitative comparisons with the best other algorithms show an improvement of the ratio between performance and computational complexity up to a factor N. The performance of the IMM algorithm reaches, for small time lags, that of an N times more complex generalized pseudo Bayes algorithm.