学习种群分布特性的演化算法研究

The loss of diversity control and blindness searching caused by over-randomness are two major challenges of evolutionary algorithms (EAs). The former refers to the lack of active control strategy of population diversity in the design of EAs. The essence of the latter is the randomness of guiding direction in the search process of population. Based on statistical learning theory, the distribution of population and evolutionary mechanism of local optimal individual are studied extensively in this project. At the same time, some characteristics of evolutionary population are also exploited, which are closely related to the diversity of population and searching directions. And then, active control of diversity and dynamic adjustment of searching directions are proposed to enhance the whole performance of EAs. The key research points of this project are listed as follows. Firstly, the relationship between the current optimal individuals and the global optimal individuals is studied. Then, the method of progressive estimation for uniform approximation of optimal population is proposed to find evolutionary mechanisms of population distribution and local optimal individuals. Secondly, the population diversity is regarded as a definite optimization target. Therefore, a single objective optimization problem is transformed into a dynamic bi-objective optimization problem. Then, we propose a new feature learning approach for the distribution of evolutionary populations. The diversity is actively controlled in real time based on the learnt features, and it will further improve the search efficiency of mainstream evolutionary computation models. Thirdly, in order to reach collaborative searching, a new strategy of dynamic adjustment of searching directions is constructed by employing self-organization mapping and dynamic covariance learning. In this strategy, the individual can dynamically learn the global and local correlation of population distribution information, and it will help to achieve the minimum direction of data association. Finally, the feasibility and effectiveness of the proposed approach are evaluated by conducting several practical applications. This application is related to optimal control of train dispatching in Urban Rail Transit. This work aims to enrich the theory and methodology of population characteristics learning based EAs, which is important to conduct stability analysis and control strategy. Meanwhile, it is also helpful to further expend the applied space of EAs. Several research findings with intellectual property are acquired while accomplishing this project. Furthermore, it also aims to obtain the academic status in the corresponding research areas in the world wide. In the implementation of the project, we plan to cultivate several young academic backbones, and further create conditions for subsequent research work.

演化算法中如何在保障个体生成的随机性和种群分布的多样性基础上，构建全局与局部搜索的动态协同机制，实现降低消耗、提高效率、增加结果可靠性的目的，不仅是理论研究的关键问题，而且是应用推广的重要问题。有鉴于此，项目组拟采用点集拓扑和线性算子理论，研究种群分布和局部最优个体的演化机理，分析演化过程中种群分布特性对多样性和搜索方向的影响因素及作用规律。在此基础上，尝试提出多样性的主动控制策略和搜索方向的动态调整方法，旨在提升演化算法的整体性能。研究工作将首先分析与设计最优种群一致逼近的渐进估计方法，以奠定整体工作的理论基础；其次，探讨基于种群演化分布特征的学习方法，及基于自组织映射和动态协方差的学习方法，实现种群多样性实时主动控制和搜索方向动态调整基础上的协同搜索；最后，研究上述改进策略在工程领域应用的可行性与有效性。研究工作对丰富演化算法的方法体系，促进稳定性分析和过程控制策略的研究具有推动作用。

演化算法中如何在保障个体生成的随机性和种群分布的多样性基础上，构建全局与局部搜索的动态协同机制，实现降低消耗、提高效率、增加结果可靠性的目的，不仅是理论研究的关键问题，而且是应用推广的重要问题。有鉴于此，首先，课题组针对现有演化算法普遍存在的搜索盲目性问题，通过借助相关的机器学习技术，挖掘与演化算法性能密切相关的种群分布特征，进而构造和设计出一些预处理机制，以引导算法的搜索行为。通过仿真实验分析，发现提出的上述机制具有良好的可行性和有效性，为后续研究打下基础。其次，针对演化算法中全局和局部搜索过程间的平衡性问题，尝试利用点集拓扑和线性算子理论，试探性地给出了最佳逼近个体对全局最优个体的一致性估计方法，并在此基础上开展了一些针对最优种群的最佳一致逼近规律的研究工作。研究中，课题组分析和挖掘了蕴含在演化种群中并与种群多样性和搜索方向密切相关的一些分布特征，进而设计出指导后续多样性控制和搜索方向的一些动态调整策略，收到理想效果。最后，针对学习种群分布特性的实现方法问题，课题组从演化数据角度入手，挖掘出一些与多样性密切相关的种群概率分布特征，以及与算子搜索方向密切相关的种群分布的数据关联特征。在此基础上，课题组以所挖掘的特征为驱动，分别构建出面向种群多样性主动控制策略的动态双目标方法，和面向搜索方向整体动态调整的双坐标系统协同搜索方法，通过仿真实验分析和一些初步的应用验证分析，上述研究思路、设计方法、仿真模型等，均表现出比较理性的有效性和潜在的应用价值。通过本项目的实施，课题组在研究期内发表论文33篇、申请发明专利21项、注册软件著作权12项，在所属方向上树立了具有一定国际影响力的学术观点。对照立项时制定的客观目标，已全面保质保量完成。同时，在此过程中，培养了青年学术骨干多名，其中3人相继可以独立从事研究工作，并先后获得国家自然科学基金支持。对照立项时制定的客观目标，已全面保质保量完成。