This paper proposes a zero-order method of nonlinear optimization using back-propagation nets, refer to as neural network nonlinear programming, or NNNLP. The primary procedure includes (1) training of a network to represent an explicit or implicit objective function, (2) examination of the feasibility of mapping function, (3) addition of training sets, (4) reduction of design space, (5) retraining of a mapping network, and (6) searching of an optimum solution. NNNLP works like a parallel multi-line search instead of one-line search in traditional optimization methods. This strategy increases the possibility of obtaining a global optimal solution and provides a totally new perspective of solving an optimization problem. Several constrained and unconstrained problems are solved by this approach and compared with the existing method. The accuracy and efficiency of this method can be improved by enhancing the computer capability and neural network architecture.
關聯:
Proceedings of 1994 International Symposium on Artificial Neural Networks,頁596-601