A trust region method for the optimization of noisy functions

The optimization of noisy or not exactly known functions is a common problem occuring in various applications as for instance in the task of experimental optimization. The traditional tool for the treatment of such problems is the method of Nelder-Mead (NM). In this paper an alternative method based on a trust region approach (TR) is ooered and compared to Nelder-Mead. On the standard collection of test functions for unconstrained optimization by Mor e et al. 6], TR performs substantially more robust than NM. If performance is measured by the number of function evaluations, TR is on the average twice as fast as NM.