Researches On Reinforcement Learning Based On TileCoding Function Approximation

Function approximation is an application of supervised learning,the main topic studied in machine learning,artificial neural networks,pattern recognition,and statistical curve fitting.In the high dimensional state space of reinforcement learning,the use of neural networks can effectively solve the dimension disaster problem.In this thesis,combined with reinforcement learning knowledge,based on the function approximation and low efficiency problem of dimension,the main research results are concluded as follows:ⅰ. The use of a simple feature extraction method of TileCoding.The advantage of this method is that the characteristic parameters and state space unrelated,function generalization ability depends on the size and shape of the receptive field.The experiment proves,receptive field influence function generalization,the performance of generalization depends on the number of features.ⅱ. The use of CMAC cerebellar neural network algorithm,combined with the reinforcement learning algorithms and neural networks.CMAC is a local approximation method,for each input output data pairs,only a small amount of connection weights need to be adjusted,so that the learning speed is very fast.Using tiling to divide the state space,to ensure fast learning speed at the same time,and reduces the amount of computation.ⅲ. Based on the above theory research,design of the car experiment.