DocumentCode :
328264
Title :
Lossless divide and conquer principle for massively parallel optimizations
Author :
Szu, Harold
Author_Institution :
Dahlgren Div., Naval Surface Warfare Center, Silver Spring, MD, USA
Volume :
1
fYear :
1993
fDate :
25-29 Oct. 1993
Firstpage :
431
Abstract :
A lossless divide-and-conquer strategy is discovered for solving a global optimization on massively parallel and distributed processors. The lossless principle which is analogous to the incoherent phenomenon in physics is expressed in terms of a vector velocity V derived from the least mean square kinetic energy E=|V|2/2. We illustrate the nonlinearity that the sum of the best traveling salesman problem (TSP) solutions in subregions is not necessarily globally the best because of the boundary resultant vectors: V=A+B have a cross interaction terms (A,B)≠0. In fact, a nearest neighbor connection at the boundary cities represents a longer overall distance than the next nearest neighbor connection between boundary cities. Then, a theorem of orthogonal division error for lossless divide-and-conquer is proved, and the orthogonal projection is constructed for solving TSP explicitly.
Keywords :
computational complexity; divide and conquer methods; mathematics computing; neural nets; parallel processing; simulated annealing; travelling salesman problems; boundary resultant vectors; distributed processors; least mean square kinetic energy; lossless divide and conquer principle; massively parallel optimizations; nearest neighbor connection; neural network; nonconvex optimisation; orthogonal division error; simulated annealing; traveling salesman problem; vector velocity; Artificial neural networks; Cities and towns; Clocks; Concurrent computing; Constraint optimization; Costs; Least squares approximation; Military computing; Nearest neighbor searches; Silver;
fLanguage :
English
Publisher :
ieee
Conference_Titel :
Neural Networks, 1993. IJCNN '93-Nagoya. Proceedings of 1993 International Joint Conference on
Print_ISBN :
0-7803-1421-2
Type :
conf
DOI :
10.1109/IJCNN.1993.713948
Filename :
713948
Link To Document :
بازگشت