A Parallel Method With Hybrid Algorithms For Mixed Integer Nonlinear Programming
Kai Zhou,Wei Wan,Xi Chen,Zhijiang Shao,Lorenz T. Biegler
DOI: https://doi.org/10.1016/B978-0-444-63234-0.50046-4
2013-01-01
Abstract:This study aims at improving the solution efficiency of Mixed Integer Nonlinear Programming (MINLP) through parallelism. Unlike most conventional parallel implementations of MINLP solvers, which utilize multi-threads to share the burden in the serial mode, the proposed method combines hybrid algorithms running on different threads. Two types of algorithms are designed in a parallel structure. One is the Quesada and Grossman's LP/NLP based branch and bound algorithm (QG); the other is Tabu Search (TS). The proposed method attempts to minimize the search space through continuous communication and exchange of intermediate results from each thread. Three kinds of information are exchanged between the two threads. First, the best solution in TS, if feasible, serves as a valid upper bound for QG. Second, new approximations which can further tighten the lower bound of QG can be generated at nodes provided by the TS. Third, strong branching in QG may fix some integer variables, which can help reduce the search space of TS. Both threads can thus benefit from the exchanged information in the hybrid method. Numerical results show that solution time can be greatly reduced for the tested MINLP. In addition, complexity analysis of the parallel approach suggests that the proposed method has the potential for superlinear speedup.