一种利用圆弧搜索的修正的信赖域方法.docVIP

  A modified trust region method with arc search PAN XIN* (Science School, Tianjin University of Technology and Education, Tianjin 300222) 5 10 15 Abstract: The main idea of the trust region method for unconstrained optimization is to find a neighborhood named “trust region” around the current point in which a quadratic model agrees with objective function, and to search a better point in the region based on the quadratic model. When the objective function is very nonlinear, the radius of trust region may be very small and too many steps should be taken to search the minimizer. In order to efficiently solve the unconstrained optimization problem in which the objective function is very nonlinear, a modified trust region framework is proposed in this paper. A procedure called “arc search” is introduced to improve the trial step, and a new strategy to adjust the trust region radius is proposed which prevents the radius being too small. The global convergence of this new method is proved in this paper under some mild assumptions. Numerical tests illustrate that this new method is more efficient than the traditional trust region method. Key words: unconstrained optimization; trust region method; arc search; numerical experiments 0 Introduction We consider the following unconstrained nonlinear optimization problem min f ( x), x?∈ R n  (1) n  20  solving the problem (1), line search methods and trust region methods are two famous classes[1][2]. The common framework of line search methods is as follows: for a given point xk , find a descent direction d k and search a suitable step-length?? k along this direction. Then a new point is generated by xk??1?? xk???? k d k (2) 25 On the other hand, trust region methods try to find a neighborhood named “trust region” around the current point xk in which a quadratic model agrees with objective function, and to search a trial step pk in the region by solving the following quadratic subproblem:  p∈R T 1 T  p B