遗传算法和神经网络在软件抗衰技术中的应用

【作者】 蒋伟；

【导师】 严悍；

【作者基本信息】 南京理工大学 ， 计算机应用技术， 2007， 硕士

【摘要】 软件衰退现象，即软件系统随时间而出现的状态退化和性能降低，乃至系统崩溃的现象，是影响系统可靠性的一个重要因素。为了减缓软件衰退所带来的危害，软件抗衰技术被提了出来。目前有两种最基本的软件抗衰策略——基于时间的抗衰策略和基于测量的抗衰策略。本文围绕基于测量的软件抗衰策略，结合遗传算法和神经网络对定期监测和收集到的系统性能参数进行分析，预测资源消耗和软件衰退的趋势，在系统负载过重时采取必要的应对措施。本文首先对遗传算法和神经网络的优缺点进行分析，然后将这两种算法结合起来，使得新的结合算法能够充分利用两者的优点，既有神经网络的学习能力和鲁棒性，又有遗传算法的全局随机搜索能力。同时，本文对基本算法做了改进，对遗传算法的步骤、编码方式、参数的选取以及适应度函数的定义按照本文的需求进行设计，在对BP网络进行训练时用多种不同的学习算法做对比，选取预测效果好的方法进行实验预测。在仿真实验中，文章给出第七天的预测值，并和实际值进行比较，说明本文算法的可行性，然后对未来的数据进行预测。最后本文对比单独用遗传算法或神经网络进行仿真实验的实验结果，表明本文算法的在预测合格率方面具有较大的优越性。更多还原

【Abstract】 A common phenomenon which called "software aging" means the software systemdegrades with time. It always makes important influence to the system. In order tocounteract software aging, a proactive technique called "software rejuvenation" has beenproposed.Now there are two most popular software rejuvenation strategies, the time-basedstrategy and the measurement-based strategy. This paper discusses measurement-basedsoftware rejuvenation strategy, combined with Genetic Algorithms and neutral networkmethods to analyze the parameters of system performance which are monitored andcolleted periodically. In conclusion, it predicts resource consumption and the trend ofsoftware aging.This paper carried on the analysis to advantages and shortcoming of the BP neuralnetwork and the genetic algorithm, then this article combines the BP neural network andgenetic algorithms to full advantages of both which makes the new algorithms both havingthe BP neural network learning capability and robustness and the strong global searchcapability from the genetic algorithms. In the same time, some improvements have beenmade in this paper, such as the "growing up" process, the coding method, the parameterand the definition of the fitness function in the genetic algorithms. And during the trainingof the BP neural network, several learning methods have been compared.Comparing the result of the seventh day giving by the experiments and the test, thealgorithm in this paper has been proved to be doable. And then this paper forecast the datain the further. At last the experiment results given both by the genetic algorithm and by theneural network have been compared, and algorithm in this paper has been proved to beefficient.更多还原