Abstract:

An adaptive modeling relevance vector machine (AM-RVM) algorithm is proposed and used for state prediction of electronic systems. Compared with traditional RVM, the proposed algorithm can select the most suitable basis function for the training data through posterior probability values before training. Then the kernel parameters are selected fast and automatically through an optimized incremental learning process. Finally, AM-RVM is used for state prediction of electronic systems after phase space reconstruction of the state parameters of the electronic system. Experimental results of chaotic time series prediction and radar transmitter state prediction indicate that AMRVM outperforms the traditional algorithm in both prediction accuracy and training speed.