几类谱共轭梯度方法理论及数值行为研究

【作者】 胡朝明；

【导师】 万中；

【作者基本信息】 中南大学 ， 应用数学， 2012， 博士

【摘要】 谱共轭梯度算法的研究迄今虽已取得较丰富成果,但如何更合理选择谱系数和共轭系数,以保持谱方法和共轭梯度方法的计算优点,仍值得深入探讨。此外,合适的线搜索技术对算法效率和建立其收敛性理论亦至关重要。针对这些问题,本文对几类新的谱共轭梯度方法的理论及数值行为开展了较广泛和深入的研究,主要研究工作如下：1、鉴于Armijo线搜索可能生成较粗糙步长,我们基于近似Wolfe条件,提出了一种修正的Armijo线搜索规则。结合新的线搜索规则,本文设计了一类求解无约束优化问题的具有充分下降性的谱PRP共轭梯度算法。数值实验表明该算法是有效的。2、研究了一类新的非单调谱共轭梯度方法。一方面,该方法通过引入混合因子,将HS方法和PRP方法结合提出了共轭系数的新的选取方式。通过合适选取谱系数,我们证明了所得搜索方向不依赖于线搜索条件恒为充分下降方向。另一方面,该方法还修正了Zhang和Hager提出的非单调线搜索规则,在更弱的假设条件下证明了全局收敛性。数值试验说明了该方法的数值计算性能优良。3、给出了一种新的求解无约束优化问题的修正谱FR共轭梯度法。由此方法产生的搜索方向是充分下降方向；通过引入一调节系数,调节谱系数和共轭系数,以保证新方法尽量兼备谱方法和共轭梯度法的优越性。结合Wolfe型线搜索,我们给出了算法的全局收敛的证明。数值实验证实了算法的有效性。4、针对一系列具有充分下降的搜索方法,我们抽象了谱共轭梯度法的一般迭代格式,给出了谱系数和共轭系数的选择范围,以保证搜索方向是充分下降方向。对基于标准Armijo线搜索的任何其他线搜索条件,建立该一般迭代格式的全局收敛性理论。更多还原

【Abstract】 Although there has existed a great contribution in the research of spectral conjugate gradient methods up to now, it is significant to further investigate how to choose a suitable spectral parameter and a conjugate parameter such that the obtained method has the advantages owned by the spectral method and the conjugate gradient method. On the other hand, a suitable line search technique is important for the numerical performance of the developed algorithm and the establishment of global convergence. For these issues, in this dissertation, we intend to conduct an extensive and deep research on some types of spectral conjugate gradient methods from the viewpoints of theory and numerical performance. The main contributions are as follows:Since it is possible that the obtained step size by the standard Armijo line search is coarse, we first employ approximate conditions of Wolfe line search to improve the Armijo line search rule. On the basis of this new line search, a new spectral PRP conjugate gradient algorithm is developed for solving unconstrained optimization problems, where the search direction is sufficiently descent. Numerical experiments show that the developed algorithm is promising.Then, a new nonmonotone spectral conjugate gradient method is proposed. By introducing a hybrid coefficient, the conjugacy parameter is determined based on the combination of PRP and HS methods. A spectral parameter is appropriately chosen such that each search direction is a sufficiently descent direction independent of the employed line search techniques. On the other hand, the nonomonotone line search technique proposed by Zhang and Hager is modified, and under more mild assumptions, the global convergence of the developed algorithm is proved. Numerical experiments are employed to demonstrate the efficiency of the algorithm.Thirdly, a new modified spectral FR conjugate gradient method is presented for solving unconstrained optimization problems. The direction generated by the method produces sufficiently descent search direction. By introducing a control parameter to regulate the spectral coefficient and conjugate coefficient, the new method is assured to have the superiority of spectral method and conjugate gradient method. Global convergence results is established for the conjugate gradient methods with a Wolfe type line search. The numerical results show that the new method is efficient for general unconstrained optimization problems.Finally, based on a series of methods with sufficiently descent direction, a generic iterative formula of spectral conjugate gradient method is summarized. To make sure that the search direction is sufficiently descent direction, the value range of spectral and conjugate parameters is presented. With any line search conditions based on standard Armijo line search, the global convergence theory is established for this generic iterative formula.更多还原