协同演化算法在大规模优化中的相关问题研究

It is urgent for us to endow an algorithm with superior learning capabilities and more accurate reasoning skills since some pending problems such as scientific computing and engineering applications become unceasingly complicated as well as larger-scale. Relying on the learning and reasoning capabilities, the algorithm can find out some hidden characteristics of a target problem and then directs the optimization process based on the characteristics. . In this project, several key issues of cooperative coevolutionary algorithm (CCEA) faced for large-scale problem's optimization are investigated by integrating CCEA with machine learning (ML) theory. (1) Firstly, we will research on the unsupervised learning theory and model of CCEA by deepening the integration of ML theory and the algorithm because the evolutionary process of population in CCEA can be taken analogous to a sampling-analysis and learning-reasoning process in ML theory. (2) Secondly, based on the unsupervised learning model of CCEA, some intelligent information of the population, which is emerged from individuals' interacting during the evolutionary process, will be used to analysis some characteristics of a large-scale problem. Furthermore, some principal components in the problem are extracted based on its characteristics. As a result, we can reduce variable dimensionality of the large-scale problem. (3) Thirdly, after reducing the variable dimensionality, we will study on various characteristics of different correlations between variables based on grouping strategy and Bayesian theory, and then propose an effective adaptive grouping strategy. (4) Fourthly, we will investigate some characteristics of fitness landscape and determine trusted domains of different variables according to their statistical properties of sub-populations' evolutionary process. In addition, further researches will be conducted to design an adaptive allocation strategy of computing resources for CCEA. Finally, we will achieve a more reliable and robust CCEA for large-scale problems optimization.. In the project, intelligence algorithms are deeply integrated with machine learning theory, which is not only conforming to the development trend of intelligence algorithms' theories and applications, but also meeting the requirements of analyzing and solving large-scale and complex problems.

科学计算及工程应用的日趋大规模与复杂化要求算法具备更好的学习推理能力,能分析并归纳出问题的隐含特征,并依此指导其求解过程。项目将协同演化算法与机器学习理论相结合，研究其在大规模优化中的相关问题。1)将协同演化算法的种群进化比拟为采样分析与学习推理过程，结合机器学习相关理论来研究算法的无监督学习理论与模型；2)以协同演化算法的机器学习模型为基础，利用个体间信息交互与种群的智能涌现，对目标问题进行特征分析，研究问题主元提取策略，实现变量降维；3)结合随机分组策略和贝叶斯理论，研究变量间不同类型相关性的外化特征,探索实用的自适应分组策略；4)通过子种群进化过程体现的统计特性，研究不同搜索区域的适应值景观特征，确定变量可信域，进而设计计算资源的自适应分配策略，以实现算法整体性能优化。.项目将智能算法与机器学习理论深度融合，顺应了演化算法理论与应用的新发展，满足了大规模复杂问题分析和求解的新需求。

协同演化算法的成功应用推动了现代科技的发展，而日趋大规模化、复杂化的科学问题也对演化算法提出了更高的要求。统计与推理能力是提升演化算法在大规模优化问题中综合性能的主要策略之一。项目按照项目计划书实施，利用统计分析方法对演化算法的搜索行为进行分析，进而研究问题适应值景观特征。通过统计结果赋予演化算法一定的推理能力，实现其在大规模函数优化问题中的自适应控制及计算资源的自适应分配。具体研究成果包括：1) 基于搜索空间离散化的种群行为统计方法。该成果以搜索空间离散化为基础，对种群的历史搜索行为进行统计。根据统计结果对搜索空间进行收缩处理，提升搜索精度。同时，帮助种群有目的地跳出局部最优和探测。2) 基于多角色行为的适应值景观分析方法。利用动态多种群机制实现个体的多角色自适应调节，进而分析不同进化阶段角色所处适应值景观的特征。根据不同适应值景观的特征，赋予不同角色个体差异化搜索策略行为，实现全局搜索能力与局部搜索能力间的互补与协同。3) 演化算法自适应控制策略。将人类多层次学习过程、生物体遗忘机制与演化算法相结合，将演化算法的进化过程从单一层次扩展至两层次，即个体层与变量层，同时根据大规模问题搜索空间适应值景观特征的不同，实现个体遗忘能力的自适应控制，进而实现个体的差异化搜索行为，进而提升算法的综合性能。4) 计算资源自适应分配。项目组将多种群机制、多角色策略应用于演化算法，以适应值景观作为参考，指导种群间个体的迁移操作，实现子种群间的协作与计算资源的合理分配。5) 多评价机制。传统演化算法中一般是将个体对应的问题函数值作为评价指标，即适应值，因而种群的整个进化过程是在这种基于适应值的评价机制下驱动进行的。项目组将基于新颖性的评价机制引入到演化算法中。在不同进化阶段，根据搜索任务的不同，自适应调节两类驱动力的权重，通过新颖性和适应值两类评价机制实现两类驱动力的融合，达到增强算法普适性的目的。