|
A fuzzy logic controller ( FLC ) has been conspicuous by its capability that it can cope with the approximating estimate and control problems under large-grain yuncertainty. Besides, since the inference and the rule-based control structure, an FLC is also good in implementing the experiences and knowledge of the human experts. However, there does not exist a systematic analysis or synthesis method which can help to design a FLC. It seems to be difficult for an engineer to decide a control rule- base in the design process. Even though a rule-base has been formed by a experimental expert, the fine-tuning of the control rules appears to be limited by the elucidation of the heuristis characteristics of the FLC. In order to solve this problem, a self-learning scheme for FLC is introduced in this dissertation. This scheme can automatically form a control rule base for an FLC. After the rule base is formed, the FLC can work based on the rule base independently without the learning scheme any more. The learning scheme is based on the reinforcement learning approach. An adaptive critic unit is used to assist in solving the credit assignment probelm. It is constructed basically based on an neuron-like adaptive element and the temporal difference predict method. Another unit, call associative search unit, search a proper control strategy automatically based on the passed experience and the reinforcements supported by other unit. It is also using an neuron-like adaptive element to accumulate the experiences. The animal learning idea is induced in this unit to change the control rule. In these units, some modifications are done to get a faster learning rate and to make the learning result being more efficient. Finally, the simulated or experimental learning results are presented to show the capability of this modified learning scheme.
|