• 主题
高级搜索  >
历史搜索 清空历史
首页> 外文学位 >New sampling distributions for evolution algorithms.
【24h】

New sampling distributions for evolution algorithms.

机译：进化算法的新采样分布。

获取原文
获取原文并翻译 | 示例
页面导航
  • 摘要
  • 著录项
  • 相似文献
  • 相关主题

摘要

Evolution algorithms are stochastic optimization methods based on evolutionary principles. They have long been used in optimization, and are gaining in popularity. They are particularly useful in high dimensional problems, or in problems where gradient methods fails.; Evolution strategies, a class of evolutionary algorithms, are stochastic searches which evolve by mutation. This work proposes a new mutation distribution for use in single objective optimization. Up to now, cost function information obtained by mutations that do not improve fitness has been discarded. In many problems, particularly when cost function calls are expensive, it is desirable to use all available information to guide the search. The new method in this work patches Gaussians of different variances together to create a sampling distribution which delivers mutations designed to direct the search away from regions where low values of fitness have been observed. Analytic results for this new method are derived on idealized problems. The method is compared with existing methods on a range of test problems, and its overall performance attributes are assessed.; A new method for multiobjective optimization is also developed. Genetic Algorithms introduce innovation into their populations by a process of bit mutation. This small scale mutation is often insufficient to successfully direct the search, unless the initial population is of sufficient quality. The new method proposed here, termed Rank Biased Sampling, uses the population to create new members, which are resampled across the entire search space from a distribution designed to favor regions which are inadequately represented by the current population. Again, this method is compared to existing methods on some standard test problems.; These new optimization methods are then applied to some real-world problems of engineering interest. The optimization routines developed in this work performed well on these applications, and provide good improvements on existing methods, as well as opening up avenues for further research.
机译：进化算法是基于进化原理的随机优化方法。它们长期以来一直用于优化，并且越来越受欢迎。它们在高维问题或梯度方法失败的问题中特别有用。进化策略是一类进化算法，是通过 mutation 进化的随机搜索。这项工作提出了一种用于单目标优化的新突变分布。到目前为止，已经放弃了通过不能改善适应性的突变获得的成本函数信息。在许多问题中，尤其是在成本函数调用昂贵的情况下，希望使用所有可用信息来指导搜索。这项工作中的新方法将不同方差的高斯修补在一起，以创建一个采样分布，该分布可提供旨在指导搜索远离观察到适合度较低的区域的突变。这种新方法的分析结果来自于理想化问题。该方法与现有方法在一系列测试问题上进行了比较，并评估了其总体性能属性。还开发了一种用于多目标优化的新方法。遗传算法通过位突变的过程将创新引入其种群。除非初始种群具有足够的质量，否则这种小规模的突变通常不足以成功地指导搜索。这里提出的新方法称为“斜体”（倾斜排序），它使用总体来创建新成员，并从整个搜索空间中重新抽样，这些分布的设计旨在偏爱当前人口不足的区域。同样，在某些标准测试问题上，将该方法与现有方法进行了比较。然后将这些新的优化方法应用于一些实际的工程兴趣问题。在这项工作中开发的优化例程在这些应用程序上表现良好，并且对现有方法进行了很好的改进，并为进一步的研究开辟了道路。

著录项

相似文献

  • 外文文献
  • 中文文献
  • 专利
获取原文

客服邮箱：kefu@zhangqiaokeyan.com

京公网安备：11010802029741号 ICP备案号：京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有

  • 客服微信

  • 服务号