A Neuro-Fuzzy Controller Based On Improved Reinforcement Learning

Fuzzy control and neural network are two important theories in artificial intelligence systems. They belong to two fields that are much different and the basic theories of them are distinct too. But they can all simulate the intelligent behaviors what humans do and resolve the automatization problems which are incertitude non-linear and complicated. The amalgamation of fuzzy inference and neural network can solve the shortages they have in the process of information transaction and control solely and construct a system more intelligent and consummate.The neuro-fuzzy control system is the result of the amalgamation of these two theories. The abilities of parallel process and leaning inject new energy to the intelligent control technique and put forward new problems at the same time.On the one hand, the using of fuzzy variables needs the node have the ability to transform between numerical values calculations and fuzzy calculations which makes the transfer functions between inputs and outputs of the neures much more complex. This is a huge challenge to the hardware implement of the neuro-fuzzy control systems. In allusion to this problem, this paper provides a new kind of neure named Unidirectional Linear Response (ULR) which can transform the different kinds of transfer functions of each layer in neuro-fuzzy networks into ULR networks. The method can simplify the style of the node in a neuro-fuzzy controller and make the hardware implement easier relatively.On the other hand, what kind of learning algorithm should be chosen in a neuro-fuzzy control system? Traditional BP algorithm can be used in many kinds of multilayer neural networks. But the membership functions in fuzzy systems have needle points where the differential coefficients do not exist. And the systems use BP algorithm may stick in a local optimal point. Genetic algorithm is another choice. But genetic algorithm needs different coding methods in different problems. The coding job may became quite complex in some cases, especially when there are may parameters. Therefore, this paper provides a kind of reinforcement learning based on Linear Search. Reinforcement Learning (RL) is a non-supervise learning method. Linear Search is a species of algorithms used in parameter optimization. Many algorithms of this kind use the method of choosing point in skip which is similar genetic algorithm. This method can avoid sticking in local optimal point in some instances. And many of these algorithms are facility and simple. We combine the reinforcement learning and Linear Search and gain a compound learning algorithm used in neuro-fuzzy system.In this paper, we will introduce the unidirectional linear response unit and the reinforcement learning based on Linear Search into the neuro-fuzzy control system and design an adaptive multilayer neuro-fuzzy controller used on non-linear system. We will present the particular structure of this ULR neuro-fuzzy controller and introduce the application of new learning method. We applied this controller to a common control problem of inverted pendulum to evaluate the effectiveness. The obtained results were compared to the fuzzy controller trained by TBP algorithm under same condition and it clearly showed that the proposed method can avoid local optimal point in some instances and has more robustness and adaptability in neuro-fuzzy control system.