基于性能指标与差分进化的高维多目标进化算法及其并行实现研究

Many-objective optimization, refer in particular to the optimization of multi-objective problems with more than three objectives, is an emergent research front in evolutionary computation. Aiming at the issue of difficult to accurately and efficiently solve many-objective optimization problems, this project will systematically and deeply investigate the design and implementation of many-objective evolutionary algorithms (MaOEAs) from the aspects of algorithm structure, recombination operator and parallel implementation. On the basis of investigating the principles and features of indicator-based MaOEAs, an effective algorithm framework will be constructed. Through exploiting the effective connecting points between differential evolution (DE) and algorithm framework and improving DE operator from multiple angles, some new MaOEAs will be designed. Afterwards, the project will investigate the mechanisms and strategies of the parallel implementation of the proposed MaOEAs by using CUDA technique. Through testing and verifying algorithm performance via different many-objective optimization problems with various complexities, this project accomplishes the objective of proposing several MaOEAs with good optimization performance and running efficiency. This project will provide new approaches and tools for solving practical many-objective optimization problems, and also offer a new thought for the design and implementation of other kinds of MaOEAs.

高维多目标优化特指目标维数大于3的多目标问题优化，是当前进化计算的前沿研究领域。针对高维多目标优化问题难以准确和高效率求解的问题，本项目从算法结构、重组算子和并行实现等角度系统深入地研究高维多目标进化算法(Many-objective Evolutionary Algorithms, MaOEAs)的设计与实现。以基于性能指标MaOEAs原理与特性分析为基础，构建有效的算法框架；挖掘差分进化与算法框架有效结合方式并从多角度对差分进化算子性能进行改进，进而设计几种新的MaOEAs；以CUDA技术为工具，研究所设计MaOEAs的并行实现机制与策略。采用不同复杂程度的高维多目标优化问题对算法性能进行测试与验证，实现提出具有优异优化性能和较高实现效率MaOEAs的项目研究目标。本项目研究成果将为实际工程中的高维多目标优化问题求解提供新方法和新工具，也为其它类型MaOEAs的设计与实现提供新思路。

工程实际中存在大量具有搜索空间广、目标个数多等特性的复杂优化问题，进化算法作为一种随机搜索策略，是解决这类问题的有效途径。本项目针对基于性能指标的多目标进化算法进行研究，并以工程实际为背景探索进化算法的应用。本项目研究内容如下：针对电力系统环境经济调度问题，研究多目标差分进化算法的应用，解决现有方法存在的不足；分析基于性能指标多目标算法的机理，包括稳定状态模式和一般状态模式的优劣、以及基于不同性能指标下的个体评价、亲代选择和环境选择等，在此基础上设计了基于自适应性能指标选择的多目标差分进化算法；为探索多目标进化算法与图像处理的结合，项目对数字图像修复问题进行了研究，提出基于方向结构分布分析和基于多方向特征能量方程的图像修复方法。本项目研究工作为后续开展基于进化计算神经网络架构搜索的图像修复奠定了基础。项目发表学术论文5篇，获授权发明专利2项，培养硕士研究生4名。