The ping-pong robotic system, which refers to the vision measurement, pattern recognition, physically modeling, machine learning and so on, has attracted much research attention in recent years. In this dissertation, the machine learning methods in the ping-pong robotic system are investigated. The main contributions of the thesis are stated as follows: Firstly, a memory-based fuzzy learning approach is proposed for trajectory prediction in the ping-pong robotic system. The experience data are stored in multiple subspaces which are obtained by dividing the input space according to the fuzzy membership functions. Due to the continuous increment of the experience data in the real game, a kernel function based update mechanism is proposed to reduce the data storage in the subspaces. The regularization algorithm is used to generalize the data subsets in different subspaces independently. Then, a series of local models for the data subsets are obtained. These local models will be used for trajectory prediction. The outputs of the local models are smoothly integrated by using a fuzzy weighted algorithm. A robust technique is introduced to ensure that the regularization algorithm is well-posed. Secondly, a parallel fuzzy learning approach is proposed for determining the hitting point where the racket attached to the ping-pong playing robot will intercept the incoming ball. A series of candidate hitting points are obtained according to the trajecotry prediction of flying ball. The nearest neighbor method is used to estimate the racket velocity for each candidate hitting point, and then the approximate acceleration of the racket is obtained. A parallel fuzzy learning system consisting of two fuzzy subsystems is used to compute the success rate for the racket acceleration. Both these subsystems will be updated online based on the feedback. A performance function of the racket acceleration and the success rate is formulated to evaluate the candidate hitting points, and then the optimal hitting point is chosen. Thirdly, an active learning approach is proposed to control the racket so that the incoming ball is returned to a desired position. Two maps that are implemented with the locally weighted regression (LWR) are used to calculate the racket’s initial parameters. The active learning approach that is based on the fuzzy cerebellar model articulation controller (CMAC) is used to calculate the racket’s adjustment parameters. The racket’s final parameters are the...
修改评论