Towards a High-Performance Implementation of the MCSFilter Optimization Algorithm

Multistart Coordinate Search Filter (MCSFilter) is an optimization method suitable to find all minimizers – both local and global – of a non convex problem, with simple bounds or more generic constraints. Like many other optimization algorithms, it may be used in industrial contexts, where execution time may be critical in order to keep a production process within safe and expected bounds. MCSFilter was first implemented in MATLAB and later in Java (which introduced a significant performance gain). In this work, a comparison is made between these two implementations and a novel one in C that aims at further performance improvements. For the comparison, the problems addressed are bound constraint, with small dimension (between 2 and 10) and multiple local and global solutions. It is possible to conclude that the average time execution for each problem is considerable smaller when using the Java and C implementations, and that the current C implementation, though not yet fully optimized, already exhibits a significant speedup.