We propose a neural network model with transient chaos. or a transient
ly chaotic neural network (TCNN) as mt approximation method for combin
atorial optimization problems, by introducing transiently chaotic dyna
mics into neural networks. Unlike conventional neural networks only wi
th point attractors, the proposed neural network has richer and more f
lexible dynamics, so that ii can be expected to have higher ability of
searching for globally optimal or near-optimal solutions. A significa
nt property of this model is that the chaotic neurodynamics is tempora
rily generated for searching and self-organizing, and eventually vanis
hes with autonomous decrease of a bifurcation parameter corresponding
to the ''temperature'' in the usual annealing process. Therefore, the
neural network gradually approaches, through the transient chaos. to a
dynamical structure similar to such conventional models as the Hopfie
ld neural network which converges to a stable equilibrium point. Since
the optimization process of the transiently chaotic neural network is
similar to simulated annealing, not in a stochastic way but in a dete
rministically chaotic way, the new method ir regarded as chaotic simul
ated annealing (CSA). Fundamental characteristics of the transiently c
haotic neurodynamics me numerically investigated with examples of a si
ngle neuron model and the Traveling Salesman Problem (TSP). Moreover,
a maintenance scheduling problem for generators in a practical power s
ystem is also analysed to verify practical efficiency of this new meth
od.