准确预测处理器性能对计算机硬件设计与改进有着重要意义。然而,处理器预测系统存在两个核心问题:预测过程中处理器内部构造复杂和不确定性以及预测结果的不可解释性。置信规则库作为一种基于IF-THEN规则的建模方法,具有一定的可解释性...准确预测处理器性能对计算机硬件设计与改进有着重要意义。然而,处理器预测系统存在两个核心问题:预测过程中处理器内部构造复杂和不确定性以及预测结果的不可解释性。置信规则库作为一种基于IF-THEN规则的建模方法,具有一定的可解释性并且可以处理复杂系统评估与预测中的不确定信息。但BRB的规则爆炸问题限制了专家知识的使用。因此,本文提出了一种基于近似置信规则库(ABRB)的处理器性能预测模型。该模型通过构建单属性BRB模型来解决规则爆炸问题,并通过基于投影协方差矩阵自适应进化策略(P-CMA-ES)算法对专家知识给定的初始参数进行优化。最后以UCI中处理器数据集为例,验证了所提方法的有效性。Accurate prediction of processor performance is important for computer hardware design and improvement. However, there are two core problems in processor prediction systems: the complexity and uncertainty of processor internals during the prediction process and the non-interpretability of the prediction results. Belief rule base (BRB), as a modelling method based on IF-THEN rules, has some interpretability and can handle uncertain information in the evaluation and prediction of complex systems. However, the rule explosion problem of BRB limits the use of expert knowledge. Therefore, this paper proposes a processor performance prediction model based on approximate belief rule base. The model solves the rule explosion problem by constructing a single-attribute BRB model and optimizes the initial parameters given by the expert knowledge by the Projection Covariance Matrix Adaptive Evolutionary Strategy (P-CMA-ES) based algorithm. Finally, the effectiveness of the proposed method is validated using the UCI mid-processor dataset as an example.展开更多
With the increment of focal elements number in discernment framework,the computation amount in Dezert-Smarandache Theory (DSmT) will exponentially go up. This has been the bottleneck problem to block the wide applicat...With the increment of focal elements number in discernment framework,the computation amount in Dezert-Smarandache Theory (DSmT) will exponentially go up. This has been the bottleneck problem to block the wide application and development of DSmT. Aiming at this difficulty,in this paper,a kind of fast approximate reasoning method in hierarchical DSmT is proposed. Presently,this method is only fit for the case that there are only singletons with assignment in hyper-power set. These singletons in hyper-power set are forced to group through bintree or tri-tree technologies. At the same time,the assignments of singletons in those different groups corresponding to each source are added up respectively,in order to realize the mapping from the refined hyper-power set to the coarsened one. And then,two sources with the coarsened hyper-power set are combined together according to classical DSm Combination rule (DSmC) and Proportional Conflict Redistribution rule No. 5 (PCR5). The fused results in coarsened framework will be saved as the connecting weights between father and children nodes. And then,all assignments of singletons in different groups will be normalized respectively. Tree depth is set,in order to decide the iterative times in hierarchical system. Finally,by comparing new method with old one from different views,the superiority of new one over old one is testified well.展开更多
An intensity-based non-rigid registration algorithm is discussed, which uses Gaussian smoothing to constrain the transformation to be smooth, and thus preserves the topology of images. In view of the insufficiency of ...An intensity-based non-rigid registration algorithm is discussed, which uses Gaussian smoothing to constrain the transformation to be smooth, and thus preserves the topology of images. In view of the insufficiency of the uniform Gaussian filtering of the deformation field, an automatic and accurate non-rigid image registration method based on B-splines approximation is proposed. The regularization strategy is adopted by using multi-level B-splines approximation to regularize the displacement fields in a coarse-to-fine manner. Moreover, it assigns the different weights to the estimated displacements according to their reliabilities. In this way, the level of regularity can be adapted locally. Experiments were performed on both synthetic and real medical images of brain, and the results show that the proposed method improves the registration accuracy and robustness.展开更多
文摘准确预测处理器性能对计算机硬件设计与改进有着重要意义。然而,处理器预测系统存在两个核心问题:预测过程中处理器内部构造复杂和不确定性以及预测结果的不可解释性。置信规则库作为一种基于IF-THEN规则的建模方法,具有一定的可解释性并且可以处理复杂系统评估与预测中的不确定信息。但BRB的规则爆炸问题限制了专家知识的使用。因此,本文提出了一种基于近似置信规则库(ABRB)的处理器性能预测模型。该模型通过构建单属性BRB模型来解决规则爆炸问题,并通过基于投影协方差矩阵自适应进化策略(P-CMA-ES)算法对专家知识给定的初始参数进行优化。最后以UCI中处理器数据集为例,验证了所提方法的有效性。Accurate prediction of processor performance is important for computer hardware design and improvement. However, there are two core problems in processor prediction systems: the complexity and uncertainty of processor internals during the prediction process and the non-interpretability of the prediction results. Belief rule base (BRB), as a modelling method based on IF-THEN rules, has some interpretability and can handle uncertain information in the evaluation and prediction of complex systems. However, the rule explosion problem of BRB limits the use of expert knowledge. Therefore, this paper proposes a processor performance prediction model based on approximate belief rule base. The model solves the rule explosion problem by constructing a single-attribute BRB model and optimizes the initial parameters given by the expert knowledge by the Projection Covariance Matrix Adaptive Evolutionary Strategy (P-CMA-ES) based algorithm. Finally, the effectiveness of the proposed method is validated using the UCI mid-processor dataset as an example.
基金Supported by the National Natural Science Foundation of China (No. 60804063)
文摘With the increment of focal elements number in discernment framework,the computation amount in Dezert-Smarandache Theory (DSmT) will exponentially go up. This has been the bottleneck problem to block the wide application and development of DSmT. Aiming at this difficulty,in this paper,a kind of fast approximate reasoning method in hierarchical DSmT is proposed. Presently,this method is only fit for the case that there are only singletons with assignment in hyper-power set. These singletons in hyper-power set are forced to group through bintree or tri-tree technologies. At the same time,the assignments of singletons in those different groups corresponding to each source are added up respectively,in order to realize the mapping from the refined hyper-power set to the coarsened one. And then,two sources with the coarsened hyper-power set are combined together according to classical DSm Combination rule (DSmC) and Proportional Conflict Redistribution rule No. 5 (PCR5). The fused results in coarsened framework will be saved as the connecting weights between father and children nodes. And then,all assignments of singletons in different groups will be normalized respectively. Tree depth is set,in order to decide the iterative times in hierarchical system. Finally,by comparing new method with old one from different views,the superiority of new one over old one is testified well.
基金Supported by National Natural Science Foundation of China (No60373061)Joint Programof National Natural Science Foundation of ChinaGeneral Administration of Civil Aviation of China (No60672168)
文摘An intensity-based non-rigid registration algorithm is discussed, which uses Gaussian smoothing to constrain the transformation to be smooth, and thus preserves the topology of images. In view of the insufficiency of the uniform Gaussian filtering of the deformation field, an automatic and accurate non-rigid image registration method based on B-splines approximation is proposed. The regularization strategy is adopted by using multi-level B-splines approximation to regularize the displacement fields in a coarse-to-fine manner. Moreover, it assigns the different weights to the estimated displacements according to their reliabilities. In this way, the level of regularity can be adapted locally. Experiments were performed on both synthetic and real medical images of brain, and the results show that the proposed method improves the registration accuracy and robustness.
