The article develops a unified approach for hard optimization problems involving data association, i.e. the assignment of elements viewed as "data" {x/sub i/}, to one of a set of classes, (C/sub j/), so as to minimize the resulting cost. The diverse problems which fit this description include data clustering, statistical classifier design to minimize probability of error, piecewise regression, structured vector quantization, as well as optimization problems in graph theory, e.g. graph partitioning. Whereas standard descent-based methods are susceptible to finding poor local optima of the cost, the suggested approach provides some potential for avoiding local optima, yet without the computational complexity of stochastic annealing. The approach we develop is based on ideas from information theory and statistical physics, and builds on the work of Rose, Gurewitz, and Fox (see IEEE Trans. on Inform. Theory, vol.38, p.1249-58, 1992) for clustering and related problems.
展开▼
机译：本文针对涉及数据关联的硬优化问题开发了一种统一的方法，即将被视为“数据” {x / sub i /}的元素分配给一组类之一（C / sub j /），以便从而最大程度地降低了成本。适合该描述的各种问题包括数据聚类，使错误概率最小化的统计分类器设计，分段回归，结构化矢量量化以及图论中的优化问题，例如图分区。尽管基于标准下降的方法容易发现较差的局部最优成本，但建议的方法为避免局部最优提供了一些潜力，但没有随机退火的计算复杂性。我们开发的方法基于信息论和统计物理学的思想，并以Rose，Gurewitz和Fox的工作为基础（请参阅IEEE Trans。on Inform。Theory，vol.38，p.1249-58，1992），聚类和相关问题。
展开▼
