基于广义回归神经网络的粒子滤波算法研究

Research on GRNN-based particle filter algorithm

针对基本粒子滤波算法存在的粒子退化问题,提出了一种基于广义回归神经网络(GRNN)的重要性样本调整的粒子滤波算法。利用广义回归神经网络优化从重要性密度函数采样的样本,将样本作为神经网络的输入,以观测值作为神经网络的目标向量,通过多次训练优化光滑因子逼近目标向量,用样本值和其周围的调整值作为训练后神经网络的输入向量,通过神经网络的输出向量指示用最优点来取代样本值。利用GRNN对样本进行调整,使得样本更接近于后验概率密度。仿真结果表明:基于广义回归神经网络的粒子滤波算法的性能在有效粒子数和均方误差参数方面优于基本粒子滤波算法,在改善滤波精度方面取得了较好的效果,验证了广义回归神经网络在粒子滤波算法中是可用的和有效的。

Aiming at the weight degeneracy phenomena in fundamental particle filter algorithm, the par- ticle filtering algorithm improving the important samples based on generalized regression neural network （GRNN） was presented. This algorithm optimizes the samples from importance density function by GRNN. The samples of particle filter are selected as the neural network input and the observed values as the neural network target vector. The smooth factors are optimized to approach the target vector through multiple trainings, and the samples and the corresponding surrounding adjusted values are regarded as the input of the neural network after training. The best optimized values are used to replace the sample indicated by the output vector indicator of neural network. The samples are adjusted through GRNN so as to be closer to the posterior probability density. Simulation results show that the GRNN-based particle filter algorithm, superior to the fundamental algorithm, can increase the effective particle number, reduce the mean square error, and improve the accuracy of the filtering performance. It is proved that this GRNN is available and effective in the particle filter algorithm.