Research On Sockt Forecasting System Based On Reinforcement Learning

As establishing nonlinear mode using artificial neural network has been widely used to economy forecasting for recently years, this paper will do some research on the topic. Using BP neural networks on stock forecasting system is a new research field and has a great evolvement. But this method is not good at stability and definition. So it make the system's self-study ability not strong, when the data quantum become bigger, the forecasting result become weaker reference value.In this paper, we provides the optimization on a reinforcement learning algorithm based on neural network. Using this method we can improve the system's stability and definition and improve the generalization ability of learning system.(l)Choose proper socket data to operate, assure the data have a definite reference value.(2)Choose the proper method to pretreatment the data, assure the data have good astringency and faster constringency rate during the BP neural networks training process, the system performance gained definite ensure.(3)Adopt the appropriate algorithm based on neural network ensemble, assure the forecasting system's stabilization, conquer the BP neural network system's instability. So that the forecasting result have higher reference value and the forecasting stock price curve has better stability.(4)Analysis the characteristic of time serial, adopt the right reinforcement learning algorithm to optimize the result of the neural network's output, choose the excellent parameter of the reinforcement learning system.All algorithms above are implemented in MATLAB, and utilizing VC++ and MATLAB realization the system, confirm excellent parameter through our test, validate the rationality of our system. Analysis the advantages of the designof our system, make sure that our design is succeed.