Interactive simulation of backpropagation and creeping-random-search learning in neural networks

Abstract

A new environment for interactive neural-network experiments on personal computers, DESIRE/NEUNET combines a readable matrix language with very fast "direct execution". Experimenters enter and screen-edit simulation programs and run immediately, without the usual annoying delays for compilation and linking. Simula tions still run much faster than Microsoft FORTRAN. Users can compose their own help screens and interactive menus. DESIRE/ NEUNET permits conventional dynamic-system simulation (up to 1000 differential equations) as well as, or in conjunction with, neural-net simulation and has been used to simulate Hopfield and bidirectional associative memories, transver sal predictors, and various competitive- learning schemes. This report exhibits the classical XOR training experiment with backpropagation and creeping-random- search learning.

Keywords

interactive simulation personal computers neural networks optimization

Get full access to this article

View all access options for this article.

References

Korn, G.A. : Interactive Simulation and Mathematics on Personal Computers, McGraw-Hill, New York, 1988.

A New Environment for Interactive Neural-network Experiments , Neural Networks, 2, 229-237, May, 1989 .

Rumelhart, D.E. , and J.L. McClelland : Parallel Distributed Processing, MIT Press, Cambridge, Massachusetts, 1986.

Andrews, M. , and G.A. Korn : Hybrid-computer Methods for Direct Functional Optimization, IEEE TC, c-24, No. 10, 958-965, October, 1975.

Kom, G.A. : Neural-network Experiments on Personal Computers and Workstations, MIT Press, Cambridge, MA, 1991 (in print).

Rastrigin, A.L. : "Random Search in Optimization Problems" (in Russian), transl. by U.S. Air Force Systems Command, foreign Technology Divsion, Wright-Patterson AFB, Ohio, 1967.

Matyas, J. : "Random Optimization" (in Russian), Automation and Control, 26, 246-253, 1965.