Abstract
Analog neural network algorithms for solving combinatorial optimization problems are analyzed on the basis of the saddle point theorem in mathematical programming theories. A generalized Hopfield network scheme is shown to be a gradient system for searching a saddle point of Lagrange functions of continuous relaxation problems of 0-1 integer programs. This derivation of neural network algorithms provides us an interpretation of deterministic annealing procedures on the basis of mathematical programming theories. Additional interpretation of the generalized Hopfield network scheme is also given as a gradient projection method in a Riemannian space. The Lagrange function is shown to be a Liapunov function for this gradient algorithm whose convergence is then guaranteed from Liapunov stability theorems.
| Original language | English |
|---|---|
| Pages (from-to) | 353-364 |
| Number of pages | 12 |
| Journal | Journal of artificial neural networks |
| Volume | 2 |
| Issue number | 4 |
| Publication status | Published - Dec 1 1995 |
Fingerprint
All Science Journal Classification (ASJC) codes
- Engineering(all)
Cite this
Mathematical programming formulations for neural combinatorial optimization algorithms. / Urahama, Kiichi.
In: Journal of artificial neural networks, Vol. 2, No. 4, 01.12.1995, p. 353-364.Research output: Contribution to journal › Article
}
TY - JOUR
T1 - Mathematical programming formulations for neural combinatorial optimization algorithms
AU - Urahama, Kiichi
PY - 1995/12/1
Y1 - 1995/12/1
N2 - Analog neural network algorithms for solving combinatorial optimization problems are analyzed on the basis of the saddle point theorem in mathematical programming theories. A generalized Hopfield network scheme is shown to be a gradient system for searching a saddle point of Lagrange functions of continuous relaxation problems of 0-1 integer programs. This derivation of neural network algorithms provides us an interpretation of deterministic annealing procedures on the basis of mathematical programming theories. Additional interpretation of the generalized Hopfield network scheme is also given as a gradient projection method in a Riemannian space. The Lagrange function is shown to be a Liapunov function for this gradient algorithm whose convergence is then guaranteed from Liapunov stability theorems.
AB - Analog neural network algorithms for solving combinatorial optimization problems are analyzed on the basis of the saddle point theorem in mathematical programming theories. A generalized Hopfield network scheme is shown to be a gradient system for searching a saddle point of Lagrange functions of continuous relaxation problems of 0-1 integer programs. This derivation of neural network algorithms provides us an interpretation of deterministic annealing procedures on the basis of mathematical programming theories. Additional interpretation of the generalized Hopfield network scheme is also given as a gradient projection method in a Riemannian space. The Lagrange function is shown to be a Liapunov function for this gradient algorithm whose convergence is then guaranteed from Liapunov stability theorems.
UR - http://www.scopus.com/inward/record.url?scp=0029462616&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=0029462616&partnerID=8YFLogxK
M3 - Article
AN - SCOPUS:0029462616
VL - 2
SP - 353
EP - 364
JO - Journal of artificial neural networks
JF - Journal of artificial neural networks
SN - 1073-5828
IS - 4
ER -