| Summary: | 博士 === 淡江大學 === 資訊工程學系 === 88 === In this thesis, we combine the advantages of the fuzzy theory, genetic algorithm, grey prediction theory and reinforcement learning to construct some tuning methods to design fuzzy control systems.
At first, A multi-objective fitness function with flexible performance evaluated items is offered to choose the appropriate parameters in the scalar-tuning mechanism such that the controlled system has a desired control performance. A multi-tuning fuzzy control system that involves fuzzy control rule tuning mechanism and dynamic scalar-tuning mechanism is proposed. A fitness function is constructed to determine an appropriate parameter set such that not only the selected fuzzy control structure has fewer fuzzy rules but also the controlled system has a good control performance.
In order to avoid the problem of growth of partitioned grids in some complex system, we used an aggregation of hyperrectangulars with different size and different positions to approximate fuzzy partitions. The parameters defining these hyperrectangulars are to be selected by using GA with a special fitness function which takes the number of fuzzy rules and the error index into account. An on-line tuning and real-time parameter tuning fuzzy control systems based on the concept of reinforcement learning are proposed such that the output of the controlled system has the desired output without knowing the mathematical model of the controlled system. A state evaluator is used to evaluate a scalar value to indicate the status of the current state, then a parameter modifier is proposed to tune the adjustable parameters. The trial-and-error tuning algorithm is withdrawn such that the fuzzy controller has the on-line tuning ability.
We combines the advantages of the grey prediction theory, fuzzy theory and genetic algorithm to design a dynamic grey prediction controller. The dynamic forecasting step is generated by a fuzzy system whose parameters are selected by GA. The dynamic grey prediction structure is proposed so that the short rise time and small overshoot of the controlled system can be take into account.
The proposed fuzzy systems with some tuning methods are illustrated by computer simulations. All the simulation results demonstrate the robustness and efficiency of the constructed fuzzy systems. The proposed tuning methods approach the desired specification of designer''s purposes.
|
|---|
