Neural network model for time series prediction by reinforcement learning

Two important issues when constructing a neural network (NN) for time series prediction: proper selection of (1) the input dimension and (2) the time delay between the inputs. These two parameters determine the structure, computing complexity and accuracy of the NN. This paper is to formulate an autonomous data-driven approach to identify a parsimonious structure for the NN so as to reduce the prediction error and enhance the modeling accuracy. The Reinforcement Learning based Dimension and Delay Estimator (RLDDE) is proposed. It involves a trial-error learning process to formulate a selection policy for designating the above-mentioned two parameters. The proposed method is evaluated by the prediction of the benchmark sunspot time series.