On-line Bayesian regression mixture model for robot model learning

Abstract

The performance of a robot system heavily relies on its model. The present paper proposes an efficient online Bayesian regression algorithm based on Gaussian Mixture Model. By using the mixture model of local Gaussian experts, the algorithm decouples global correlation of data and achieves linear computational cost to the size of the local model set. The proposed algorithm also realizes on-line implementation. To manage the size of local model on-the-fly, a strategy of adding and pruning local model based on a probabilistic criteria is proposed. Additionally, a forgetting strategy to treat outliers and non-stationary system is suggested. In the end, the algorithm achieved comparable results to other on-line regression algorithms.