A comparison of methods for traversing regions of non-convexity in optimization problems (Q2192582)

The authors consider two algorithms for non-linear optimization which can be viewed as following a curvilinear search along an approximation to the continuous steepest descent path (CSDP). The main aim of the work is to investigate the effectiveness of a curvilinear search along a CSDP as a means of escaping from regions where the objective function is non-convex. In some numerical examples the performance of these two techniques is compared with other ODE-based and trust-region methods. The obtained numerical results show that this approach is promising as a way of efficient traversing regions of non-convexity of the objective function.