Previous 
Up 
Next
Though automatic control is indispensable in today's industrial plants, there are still many difficulties in determining parameters which is needed by conventional control methods for driving more complex mechanical system.
New controller techniques based on neural network has been discussed for several years, because of its potential to implement a multi-input multi-output non-linear mapping which can be improved through learning. Two sorts of such controllers have been relatively successful. One is using a neuro-fuzzy approach and the other is using backpropagation learning, first for building a network model of the process and then to use this for training the controller network.
In this research we investigate ways to realise the fuzzy method using
a Bayesian confidence propagation neural network of the SANS type, and
then develop the performance of the system using reinforcement
learning. Our experimental task was set to balance an inverted
pendulum. This sort of task is a standard problem for neural control
research because it is simple to simulate, requiring little
computational effort, and yet its solution is non-trivial. We use two
simulators developed within the project: ANSim
and
SAROS
in the parallel processing environment SIMON
.
The angle of the pendulum from the vertical and the pendulum's angular velocity are represented using fuzzy coding. Then a neuro-fuzzy controller determines a proper control action from such actions as right-small, left-large etc. according to the state. Correct actions are rewarded and erroneous ones are punished. The behaviours of the pendulum are compared with the performance achieved by using a fuzzy controller.
Some points have been cleared out in this study. The binary signal transmission between neurons forces all connection weights in the network to be set properly when the fuzzy logic is implemented in the same way as the fuzzy controller. In addition, the neuro-fuzzy controller must be implemented with a smaller time step than the fuzzy controller because the output values are binary and they can represent graded values only statistically.
Next