UKF训练RMLP网络的快速算法

A Fast Unscented Kalman Filter in Training RMLP Networks

UKF训练PMLP的优点是无需计算动态偏导数,但UKF面临计算量大的问题,降低计算量是UKF的研究重点之一。本文提出反向遍历sigma点的方法,可以显著降低UKF有2N_w+1或N_w+2个sigma点两种情况下的计算量。另外,对2N_w+1个sigma点的情况,本文提出解耦合方法,能有效降低存储协方差阵所需空间。综合反向遍历方法和解耦合方法,本文给出UKF训练RMLP的快速算法。在动态系统辨识的仿真中,本文方法节省近80%的计算量和近90%的存储空间。

As a derivative-free learning method of RMLP networks, unscented Kalman filter（UKF） faces the difficulty of heavy calculation. In this paper, the reverse scan of sigma-points is firstly proposed. It takes the advantage of the layered structure of RMLP and the particularity of sigma-points, and efficiently reduces the calculation quantum whenever UKF has 2Nw ＋ 1 or Nw ＋ 2 sigma-points. The decoupling method is secondly proposed, which lowers the storing space when UKF has 2Nw ＋ 1 sigma-points. Combining these two methods, a final fast UKF is achieved for training RMLP. In dynamic system identification simulations ,our method saves almost 80% learning time and 90% storing space.