基于多模态群集智能的进化集成学习研究

Ensemble Learning is a top research topic in the machine learning area. Generating "good and diverse" individual learner is the key point to ensure the success of ensemble learning. In this project, this problem is transferred into multimodal multi-objective optimization problem creatively and we will focus on the theory analysis and algorithms design on evolutionary ensemble learning based on swarm intelligence. First, under the multimodal single objective condition, we will study on the diversity evaluation criteria, multimodal searching and maintaining strategies, and propose the multimodal single objective evolutionary ensemble learning algorithm with homogeneous individual learners. Then, we will treat it as a multimodal multi-objective problem and study on the diversity evaluation criteria, niching strategy, multimodal Pareto-sets searching and selection strategies to construct the multimodal multi-objective evolutionary ensemble learning algorithm with homogeneous individual learners. At last, for the heterogeneous individual learners, we will study on the diversity evaluation based on the input-output relationship, and propose the evolutionary ensemble learning algorithm based on the co-evolutionary swarm intelligence with heterogeneous individual learners. Generally speaking, the purpose of this project is to build a new evolutionary ensemble learning algorithm framework based on swarm intelligence to improve the accuracy, generalization performance and computing efficiency of ensemble learning.

集成学习是机器学习中的研究热点之一，如何产生“好而不同”的个体学习器是保证集成学习性能的关键问题。本项目利用群集智能的优势，创新性地将该问题转化为多模态多目标优化问题进行解决，开展基于多模态群集智能的进化集成学习基础理论和新型算法研究。首先在多模态单目标条件下，建立多样性度量标准，确定多模态搜索及保持策略，求解同质个体学习器生成问题，构建多模态单目标进化集成学习算法；在此基础上，研究多目标情况下，多模态同质个体学习器多样性测度、小生境保持策略、多模态帕累托解集搜索及优化选择等关键问题，构建多模态多目标进化集成学习算法；最后针对异质个体学习器的生成优化问题，研究学习器输入-输出关系的多样性度量方法，建立基于多种群协同进化的异质个体学习器进化集成学习算法。本项目研究将建立基于群集智能的新型进化集成学习算法框架，综合提升集成学习的准确性、泛化性能与计算效率。

集成学习通过集成多个具有一定差异性的个体学习器来获得具有较好泛化性能的集成学习器，其在机器学习领域中具有重要的研究意义。针对不同个体学习器在组合优化中表现出典型的多模态多目标特性，本项目开展基于多模态群体智能的进化集成学习基础理论和新型算法研究。首先，建立学习器多样性衡量指标，开发基于小生境的多模态搜索及保持策略，求解单目标场景下的多模态学习器集成问题，构建多模态单目标进化集成学习算法；在此基础上，开展多模态帕累托解集搜索及优化选择等关键技术研究，构建多模态多目标进化集成学习算法，求解性能优异而结构不同的个体学习器来提高集成学习器的性能；最终应用所设计的算法解决了能源优化、生物医学、光伏模型参数识别、特征选择以及车间调度等各方面实际应用问题。本项目研究成果为进化集成学习的求解提供了新型解决方案。