Adaptive Memory Programming for Constrained Global Optimization

The problem of finding a global optimum of a constrained multimodal function has been the subject of intensive study in recent years. Several effective global optimization algorithms for constrained problems have been developed; among them, the multi-start procedures discussed in Ugray et al. [1] are the most effective. We present some new multi-start methods based on the framework of adaptive memory programming (AMP), which involve memory structures that are superimposed on a local optimizer. Computational comparisons involving widely used gradient-based local solvers, such as Conopt and OQNLP, are performed on a testbed of 41 problems that have been used to calibrate the performance of such methods. Our tests indicate that the new AMP procedures are competitive with the best performing existing ones.