Summary: | 碩士 === 淡江大學 === 機械工程學系 === 85 === In this paper,two neural networks are applied to solve
optimization problems.One is Backpropagation network, the other
is Hopfield network.The purpose ofthe Backpropagation network
method is to recognize the constrained boundary andfeasibility .
First step is to train a boundary network,a
Backpropagationnetwork is to recognize the points along the
constrained boundary and itsfeasibility. Second step is using a
global search technique to find optimumpoint from the most
possible points. One can apply the similar process to thefuzzy
constrained problem and eventually a fuzzy transformation can
beconstructed. Zooming and moving center technique combined with
a controlledrandom search can generate the optimum solution.The
equations of motion for theHopfield network with symmetric
connections always lead to a convergence tostable state. It
means the system can minimize the function. The
optimizationproblem is mapped onto the neural network in such a
way that the networkconfiguration corresponds to possible
solutions to the problem. Similarly,themethod is applied to the
problem which the constrained functions contain fuzzyinformation
under some assumptions.When constrained functions are not
explicitand analytic but only a bunch of experimental data, the
Backpropagation networkcan obviously serve as a useful vehicle
of optimizing such practicalengineering problems. The further
research is to speed up the networkconvergence. In Hopfield
network, it takes a short time to find optimum. Themethod can
only apply to the problems which constrained functions are
explicit.The further research is to improve the network in
global sense, andconveniently and correctly deciding the
parameters of synapses and currents isrequired for dealing with
the general and practical engineering designs.
|