求解一般界约束优化问题的积极集信赖域方法

本文提出了一种求解一般界约束优化问题的新方法. 每步迭代分为两个阶段. 在第一阶段, 从 当前迭代点x_k 出发, 沿着经过仿射变换后的梯度步, 得到试探点x_k¹, 记录下它的积极集. 这里用到的仿射变换矩阵不仅依赖于变量到边界的距离, 还依赖于当前迭代点的梯度以及该步迭代中的信赖域半 径. 在第二阶段, 从x_k¹ 出发, 通过在积极约束的零空间里面求解一个信赖域子问题得到新的试探点. 然后判断是否接受这个试探点作为下一个迭代点. 文中证明了算法的全局收敛性, 并且迭代点列的每 个聚点都是一阶稳定点. 文中还对国际著名的CUTEr 算例库中所有的界约束优化问题进行了测试. 数值结果表明我们的方法是有效的, 并且可以与L-BFGS-B 方法相媲美.

An active set trust region method for general bound constrained optimization

We study a new active set trust region method for general bound constrained optimization problems. Each iteration is divided into two stages. At the first stage, from x k we take a scaled gradient step to obtain a trial point x 1 k at which we record its active constraints. Here the affine scaling matrix is dependent on the distances between x k and its boundaries, its gradient and also the trust region radius. At the second stage, a trial step is taken from x 1 k by solving a QP subproblem with an ellipsoi...