基于量子学习和调节网络的多目标聚类方法研究

With the rapid development of information acquisition and information processing technology, it has been a hot topic that how to achieve high-efficiency complex data perception and recognition. On account of the problems that the deficiency of structure approximation and weak learning ability when evolutionary algorithms dealing with complicated distribution and large sample data clustering. With reference to neural regulation mechanism, this subject constructs the regulation network model with the ability of noise accommodation, generalization and keeping memory. To solve the structure characteristic of high dimension and complexity in regulation network, this subject introduces quantum parallel computation into multiobjective optimization, and forms quantum-inspired multiobjective learning algorithm, and then optimizes network structure and parameters followed by high-efficiency learning. In this subject, clustering is regarded as an optimization problem, and in order to overcome the bottleneck problems that difficulty in determining clustering parameters, the requirement of sole data distribution and bad noise immunity in single objective optimization to process complicated data clustering, multiobjective optimization framework is adopted in this subject to strengthen the robustness and universality of clustering methods. This subject forms multiobjective clustering algorithm based on quantum-inspired learning and regulation network computation model, and analyzes and solves the critical issues which lie in resolving complicated distribution and large sample data clustering through new model and learning algorithm. The critical issues consist of the design of quantum-inspired parallel strategy and clustering validity criteria, the construction of multiobjective learning framework and the analysis of computational complexity. The effectiveness and expandability of the network computation model is verified with challenging problems, the segmentation of remote sensing images and change detection.

随着信息获取与处理技术的飞速发展，如何实现高效的复杂数据感知与识别成为热点。本课题针对进化算法在处理复杂分布大样本数据聚类时缺乏结构逼近和学习能力弱，从而借鉴神经调节机理，构造具有容噪、泛化和记忆能力的调节网络模型；将量子并行计算机理引入到多目标进化，为解决调节网络高维复杂的结构特点，建立量子多目标学习算法，优化网络结构和参数，实现高效学习；将聚类看作为一个优化过程，克服单目标优化处理复杂据聚类时存在的聚类参数难以确定、要求数据分布单一、抗噪能力差等瓶颈问题，采用多目标优化框架，增强聚类方法的鲁棒性和普适性，建立基于量子学习和调节网络计算模型的多目标聚类算法，分析并解决新模型与学习算法在求解复杂分布和大样本数据聚类中的关键问题，包括量子并行策略设计及聚类有效准侧设计，多目标学习框架构造，计算复杂度分析等；用具有挑战性的遥感图像分割和变化检测问题，来验证该网络计算模型的有效性与可扩展性。

针对进化算法在处理复杂数据聚类时缺乏结构逼近和学习能力弱，模拟生物神经调节网络具有的自适应学习、识别与记忆等功能，建立了调节网络计算模型；针对聚类分析时受初始值敏感影响、聚类参数难以确定、要求数据分布单一和扩展性弱等问题，建立了量子多目标聚类算法，实现了自动聚类；并将以上理论成果成功应用于了网络信息处理和图像理解中的关键问题中。研究成果发表论文32篇，其中SCI检索的国际期刊论文21篇，包括SCI II区的国际刊物7篇，申报国家发明专14项（其中授权5项），培养博士和硕士7人；受IEEE Fellow，IEEE CIS主席，伯明翰大学首席教授Xin Yao邀请，于2014年3月到2015年3月进行了为期1年的交流访问工作；2013年入选教育部新世纪优秀人才支持计划，2014年入选陕西省青年科技新星。