CAJ | 학술논문

对于复杂输入的贝叶斯网络，精确推理时间较长。文中针对贝叶斯网络精确推理中的团树传播算法，提出了一种基于CPU-GPU异构计算平台的并行化方法。首先研究团节点间信念势更新方式，提出了节点级并行化方法加速更新过程；其次，提出了利用计算复杂度的优先级队列方法，通过拓扑级并行化加速全局推理过程；最后，通过输入不同团树结构-线性结构、两分支二叉树结构和完全二叉树结构验证算法加速效果。实验结果表明，节点级并行化方法对线性结构有明显加速效果，拓扑级并行化对两分支二叉树和满二叉树结构有明显加速效果。
대우복잡수입적패협사망락，정학추리시간교장。문중침대패협사망락정학추리중적단수전파산법，제출료일충기우CPU-GPU이구계산평태적병행화방법。수선연구단절점간신념세경신방식，제출료절점급병행화방법가속경신과정；기차，제출료이용계산복잡도적우선급대렬방법，통과탁복급병행화가속전국추리과정；최후，통과수입불동단수결구-선성결구、량분지이차수결구화완전이차수결구험증산법가속효과。실험결과표명，절점급병행화방법대선성결구유명현가속효과，탁복급병행화대량분지이차수화만이차수결구유명현가속효과。
For Bayesian network with complex input,the inferring is time-consuming. Aiming at the junction tree propagation algorithm of Bayesian network' exact inference,a parallel method is proposed based on the CPU-GPU heterogeneous computing platform. Firstly, the updating method between junction tree nodes is investigated,and node-level parallelism method is proposed to accelerate the upda-ting. Secondly,the priority queue method is presented according to the computational complexity to achieve topological-level parallelism to accelerate the global inference process. Finally,speedup of the proposed method is verified on various input junction trees including lin-ear structure,two branches of binary tree structure and complete binary tree structure. The experimental results indicate that the node-level parallelism method can significantly accelerate the linear structure,and topological-level parallelism is effective for the two branches of binary tree and complete binary tree.