| Summary: | 碩士 === 國立清華大學 === 動力機械工程學系 === 85 === When a look-up table is added to a control system, the accuracy ofthe responses and the number of cells (or required memory) areconflicting, in general. To improve the accuracy, refining the cells maybe a better way, but it will result in more memory required. Most on-linelearning controllers need powerful computer or heavy computingdemand to executing their algorithms. In this thesis, an on-line learningcontroller is proposed using a hierarchically cell-to-cell mapping method.It employs the modified random search algorithm to improve theaccuracy and to eliminate the oscillation and chattering. The generatedcontrol table is hierarchically refined and needs only limited knowledge of plant. The presented hierarchical cells are refined according to the userdefined performance index. This thus reduces the required memorysignificantly, while the control accuracy is still achieved. To update thecontrol table, an algorithm with rule base is implemented. This approachconsumes less computational time and needs less hardwareimplementations due to simple structure. In each cell the optimal controleffort is tuned by the random search algorithm. The learning rated ofproposed method is approximately exceeding 200 reinforcements persecond which can avoid too much overshoot and thus prevent fromchattering. An experiment using a simple transmission system and asimulation using an inverted pendulum are conducted to demonstrate theperformances of the developed method.
|
|---|
