基于相似历史信息迁移学习的进化优化框架

Evolutionary Optimization Framework Based on Transfer Learning of Similar Historical Information

现有进化算法大都从问题的零初始信息开始搜索最优解,没有利用先前解决相似问题时获得的历史信息,在一定程度上浪费了计算资源.将迁移学习的思想扩展到进化优化领域,本文研究一种基于相似历史信息迁移学习的进化优化框架.从已解决问题的模型库中找到与新问题匹配的历史问题,将历史问题对应的知识迁移到新问题的求解过程中,以提高种群的搜索效率.首先,定义一种基于多分布估计的最大均值差异指标,用来评价新问题与历史模型之间的匹配程度;接着,将相匹配的历史问题的知识迁移到新问题中,给出一种基于模型匹配程度的进化种群初始化策略,以加快算法的搜索速度;然后,给出一种基于迭代聚类的代表个体保存策略,保留求解过程中产生的优势信息,用于更新历史模型库;最后,将自适应骨干粒子群优化算法嵌入到所提框架,给出一种基于相似历史信息迁移学习的骨干粒子群优化算法.针对多个改进的典型测试函数,实验结果表明,所提迁移策略可以加速粒子群的搜索过程,显著提高算法的收敛速度和搜索效率.

Most of existing evolutionary algorithms search optimal solutions from zero initial information of a given problem.Because of the lacking a mechanism to use historical information,this must waste computing resources to some extent when solving a problem similar to old ones.This paper extends the idea of transfer learning to the field of evolutionary optimization,and studies a new evolutionary optimization framework based the transfer learning of similar historical information.To improve the search efficiency of the population,the proposed framework finds out the history problem from the model library,which matches current new problem,and transfers the knowledge of the historical problem into optimization process of the new problem.First,a maximum mean discrepancy indicator based on multi-distribution estimation is defined to evaluate the matching degree between a new problem and historical models.Secondly,the knowledge of matched historical problem is transferred into the new problem,and a new initialization strategy of the population based on matching degree is given to accelerate the search speed of the algorithm.Next,a preservation strategy based on iterative clustering is presented to save good information generated during the evolutionary process,for updating the model library.Finally,embedding an adaptive bare-bones particle swarm optimization(PSO)into the proposed framework,a bare-bones PSO algorithm based on the transfer learning of similar historical information is presented.Testing on several improved typical functions,experimental results show that the proposed transfer strategy accelerates the search process of the particle swarm,and significantly improve the convergence speed and the search efficiency of the proposed PSO algorithm.