The leaky-integrator recurrent neural dynamics,
the state space search algorithm and their applictions
摘要
We study the dynamics of Leaky-integrator recurrent neural network.
Our results show that there exists at least one equilibrium point
and the set of solutions of the dynamical system is a positive
invariant and attractive set. By discretizing this dynamical system,
a state space search algorithm of the discrete-time recurrent neural
network is proposed. By searching in the neighborhood of the target
trajectory in the state space, the algorithm performs nonlinear
optimization learning process and provides the best feasible
solution for the nonlinear optimization problem. The convergence
analysis shows that the network convergence to the desired solution
is guaranteed, and the stability properties are discussed. The method
offers an ideal setting to carry out the recurrent neural network
approach to chaotic cases of data compression. It can also be applied
to solve other nonlinear least square problems for global optimization
problems, especially for power regression models.
∼邵沙麗教授演講摘要∼