动态贝叶斯网络一种自适应的局部抽样粒子滤波算法

Adaptive particle filtering for dynamic Bayesian networks inference based on local sample method

针对传统自适应粒子滤波(APF)对于动态贝叶斯网络推理中高维的问题,提出动态贝叶斯网络一种自适应的局部抽样粒子滤波算法(LSAPF)。LSAPF算法将BK算法分团的思想引入到粒子抽样中,利用策略相关性和局部模型的弱交互性为指导对动态贝叶斯网络进行分割,以降低抽样规模和抽样的状态空间;进而对局部模型用自适应粒子滤波算法进行近似推理,并以粒子的因式积形式近似系统的状态信度。实验结果表明,该算法能很好地兼顾推理精度和推理时间,其性能优于普通PF算法;与APF算法相比,在不增加推理误差的情况下推理时间也有较大的提高。

This paper proposed a local sample adaptive particle filtering （LSAPF）to resolve the high dimension problem of traditional adaptive particle filter for dynamic Bayesian networks inference.Based on particle sampling method,introduced the LSAPF algorithm the clusters idea of BK algorithm to.Acparated dynamic Bayesian networks under the guidance of strategic relevance and weak interaction of local model,so reduced the sampling model scale and state-space of sampling.Then used the APF algorithm to execute approximate inference for local model,represented the reliable state as a factor plot of particles.The experiment result shows that the LSAPF algorithm can make good choice between inference precision and inference time,and performs better than traditional particle filter algorithms.Especially compared with APF algorithm,the algorithm has greatly improved on inference efficiency without increasing inference error.