A Framework for Parallel Nonlinear Optimization by Partitioning Localized Constraints
We present a novel parallel framework for solving large-scale continuous nonlinear optimization problems based on constraint partitioning. The framework distributes constraints and variables to parallel processors and uses an existing solver to handle the partitioned subproblems. In contrast to most previous decomposition methods that require either separability or convexity of constraints, our approach is based on a new constraint partitioning theory and can handle nonconvex problems with inseparable global constraints. We also propose a hypergraph partitioning method to recognize the problem structure. Experimental results show that the proposed parallel algorithm can efficiently solve some difficult test cases.
- Research Organization:
- Washington Univ., St. Louis, MO (United States)
- Sponsoring Organization:
- USDOE
- DOE Contract Number:
- FG02-06ER25737
- OSTI ID:
- 1123918
- Report Number(s):
- DOE/ER-25737
- Resource Relation:
- Conference: International Symposium on Parallel Architectures, Algorithms and Programming, Hefei (China), 2008
- Country of Publication:
- United States
- Language:
- English
Similar Records
MO-G-304-04: Generating Well-Dispersed Representations of the Pareto Front for Multi-Criteria Optimization in Radiation Treatment Planning
LDRD final report on massively-parallel linear programming : the parPCx system.