As the recurrent neural network exhibits the excellent dynamic processing ability, a dynamic feedback control strategy using recurrent neuro-contro l is proposed to the application on the balance control of the inverted pendulum. Because the conventional error backpropagation methods for the training can not be used in the optimal design here due to that the only feedback evaluating performance is the failure signal, the extended ...·; ¶†-ES for the unsupervising learning of the control parameter is presented in this paper. Meanwhile, the stabilisation of the controlled system is guaranteed during the extended ...·; ¶†-ES learning phase using the constraints optimisation. Simulation results have shown that training eYciency of the extended ...·; ¶†-ES is better than the traditional ...·; ¶†-ES. It is also shown that the recurrent neuro-control for the dynamic system possesses excellent performance compared with the MLP neuro-control with the fewer feedback signals.
[1]
Lawrence Davis,et al.
Training Feedforward Neural Networks Using Genetic Algorithms
,
1989,
IJCAI.
[2]
L. Darrell Whitley,et al.
Genetic algorithms and neural networks: optimizing connections and connectivity
,
1990,
Parallel Comput..
[3]
Dragan Obradovic,et al.
On-line training of recurrent neural networks with continuous topology adaptation
,
1996,
IEEE Trans. Neural Networks.
[4]
Richard S. Sutton,et al.
Neuronlike adaptive elements that can solve difficult learning control problems
,
1983,
IEEE Transactions on Systems, Man, and Cybernetics.