遗传聚类算法的系统性改善策略

Clustering is inherently a highly challenging research problem, and genetic algorithm has become an important tool to approach such a problem. However, existing genetic algorithm based clustering methods suffered from a variety of inadequacies, such as difficulty of setting appropriate parameter values, inefficiency, premature convergence and difficulty in delivering reliable results, etc. This project studies a series of methods, principles and techniques, aiming at systematically improving genetic clustering algorithms. Specifically, the project will focus upon four key issues which greatly impact performance of the algorithm to study effective solutions. Firstly, for the issue of parameter setting, the project proposes to study effective methods by integrating the adaptive and self-adaptive parameter setting techniques. Secondly, for the issue of convergence speed, we propose multi-local search based hybrid genetic clustering algorithm. Thirdly, for the issue of premature convergence, the project devises a robust population diversity index based adaptive niching methods which will be integrated into the genetic algorithm to effectively explore complex, multi-model search space of the partitional clustering problems. Finally, for the issue of results' reliability, we propose to determine an appropriate weighted validity function which will be used to efficiently deliver reliable clustering results. The project will achieve efficient, effective and reliable automatic clustering which can dynamically adjust parameters of the algorithm to suit different partitional clustering problems. The research findings can be widely applied in scientific research and engineering design, therefore having considerable theoretical significance and application values.

聚类分析是一个极具挑战性的研究课题，遗传算法已成为研究该课题的重要方法。然而，现有基于遗传算法的聚类分析存在参数设置难、速度慢、局部收敛以及难以快速可靠聚类等诸多不足。本项目致力于研究一系列方法、原理和技术，旨在系统性、整体性改善遗传聚类算法。具体研究将从限制现有算法性能的四个关键问题展开。首先，结合演化自适应和自适应参数设置技术来研究有效的参数设置方法，以解决参数设置难问题。其次，提出基于多局部搜索的遗传聚类分析，来提升算法速度。再次，设计基于健壮群体多样性指标的自适应小生境技术，用于结合遗传算法对聚类问题的复杂、多模态解空间进行有效搜索，防止局部收敛。最后，研究确定一个合适的权和聚类函数，用以快速得到可靠聚类结果。本项目将实现快速、有效、可靠且具参数自调整能力的自动聚类分析。研究成果可广泛应用于科学研究和工程设计，具有重要的理论意义和应用价值。

本项目根据计划研究了改善遗传算法和数据聚类的相关理论和方法。具体工作包括：1、研究多局部搜索方法，实现基于多局部搜索操作的遗传算法；2、研究有效群体多样性指标并设计自适应小生境方法；3、研究合适权和聚类函数用以有效遗传聚类分析；4、结合演化自适应和自适应控制机制的参数设置；5、遗传算法和聚类分析在生物信息学、模式识别、安全、信息隐藏等领域的应用。期间邀请了国外专家交流访问2人次。培养硕士生5名。在IEEE Transactions等期刊和会议上发表论文11篇，包括2篇领域内的顶级、中科院SCI 1区期刊。申请发明专利6项。