关于选择性集成学习框架的拓展性研究

Selective ensemble learning deals with the selection of a model subset from the original ensemble so as to improve its efficiency and predictive performance. It is a NP-complete problem. In this project, we will carry out extended research of the selective ensemble learning framework, combined with the knowledge in the field of machine learning and pattern recognition, and based on our existing research results. Specifically, we will propose a novel selective ensemble learning method based on Greedy Randomized Adaptive Search Procedure (GRASP); we will carry out investigation of selective ensemble learning by considering diversity and accuracy simultaneously; we will carry through the design of selective ensemble learning method based on reinforcement learning; finally, we will construct a novel ensemble system based on Extreme Learning Machine (ELM), which will naturally inherit various advantages of ELM. We will carry out further research on a variety of selective ensemble learning methods on the basis of the ELM ensemble, and analyze new characteristics, performances, advantages and disadvantages of various methods in ELM ensemble. And the effectiveness of various methods will be validated through lots of experiments on the benchmark datasets, so that the existing selective ensemble learning models will be further promoted in both their performance and efficiency, in order to better meet the demand of practical applications, and make new contributions to the field of ensemble learning.

选择性集成学习研究的问题为：怎样从原始集成中选择一个模型子集，以实现对原集成在效率和预测性能上的改进。这是一个NP完全问题。本项目拟结合机器学习和模式识别领域的知识，并基于我们已有的研究成果，开展关于选择性集成学习框架的拓展性研究。具体而言，拟基于贪婪随机自适应搜索（GRASP）算法设计提出一种新型的选择性集成学习方法；开展旨在同时考虑集成多样性和精确度因素的选择性集成学习研究；基于强化学习技术设计选择性集成学习方法；最后，将基于极限学习机（ELM）构建一种新颖的神经网络集成系统，该系统将很自然地继承极限学习机的各种优点。项目以该系统为基础进一步研究多种选择性集成学习方法，分析与探讨在ELM集成中，各种方法新的特点、性能表现、优势与不足，并通过大量的基准实验验证各种方法的有效性，实现对已有的选择性集成学习模型在推广性能和效率两方面的进一步改进，更好地满足实际应用需求。

我们在项目中首先对选择性集成学习框架进行了拓展性研究，包括：选择性集成学习在模式分类、时间序列预测、与信用风险评估领域的研究，设计提出了反向降低误差集成剪枝算法、基于分支限界法的层次式并行集成选择算法等多种新颖且高效的集成学习算法。其次，我们分别开展了对深度学习、迁移学习在模式分类与时间序列预测领域的研究。第三，我们实施了增量学习算法在时间序列预测领域的研究。最后，我们分别开展了关于经典神经网络以及将集成学习方法应用于改进经典神经网络性能的研究、以及神经网络情感学习算法的研究。通过本项目的研究，我们成功地实现了对于计算机实现模式分类、时间序列预测、以及信用风险评估能力的有效提高，并且为相关基础理论做出了应有的贡献。