A Study Of Flexible Structure Control Method Based On Fuzzy Reinforcement Learning

With the development of deep space detection and radio astronomy, the large radio telescope is widely used. The new advanced radio telescope's diameter is also becoming bigger so as to get better performance.But given all these disadvantages such as structural flexibility and external disturbances, the antenna control system is requested to keep a high pointing accuracy, high tracking accuracy and more robust.This thesis takes large antenna control problem as the research background and makes reinforcement learning in artificial intelligence filed be flexible structure controller to ensure a certain tracking accuracy and to suppress the flexible vibration simultaneously.But traditional reinforcement learning method is suitable for small discrete state or small discrete action learning task, not for the problem that has large or continuous space. So in order to solve this problem, this paper brings fuzzy inference with widely approximation as approximating methods into reinforcement learning, as a result, the fuzzy reinforcement learning system can handle large space problem well. At the same time, control of flexible structure based on fuzzy reinforcement learning is verified by applying it to flexible model with continuous state space. So the main work of this paper is summarized as follows:i.In allusion to the curse of dimensionality and discrete action policy that classis Q-iteration algorithms based on Lookup-table when deal with continuous space task, this paper proposes a reinforcement learning algorithm based on type-1 fuzzy reasoning. This kind of method firstly partitions the state space into fuzzy sets; then make up fuzzy rules that take state as rule's input and take action and its Q value as rule's output, by fuzzy inference part, the continuous action can be got. Finally we update the parameters online by reinforcement learning TD-error and the gradient method.ii.By considering the reinforcement learning algorithm that based on type-1 fuzzy logic shows robust limitations,this papper proposes a new reinforcement learning algorithm based on type-2 fuzzy inference. Type-2 fuzzy system is built on type-2 fuzzy sets. While the type-2 fuzzy, the shape of whose membership functions have three-dimensional character, can describe the uncertainties in reality more appropriately. Compared to the traditional type-1 fuzzy system, type-2 fuzzy system has a stronger ability to deal with theuncertainty of system. Meanwhile for the reason that the present of interval type-2 fuzzy sets greatly simplifies the computation between the general fuzzy sets. This thesis designs a reinforcement learning method based on interval type-2 fuzzy inference. It can not only deal with continuous problems well, but also more robust to noise.iii.The two algorithms proposed in this paper are verified by simulation control of flexible model. Simulation result shows that the control of flexible structure based on Type-1 fuzzy reinforcement learning can ensure a certain tracking accuracy and suppress the flexible vibration well simultaneously compared to traditional reinforcement learning algorithms; Simulation with external disturbance also shows that the control of flexible structure based on Type-2 fuzzy reinforcement learning is more robust than the former.