A Constructive Spiking Neural Network for （建设性的飙升神经网络）.pdfVIP

A Constructive Spiking Neural Network for Reinforcement Learning in Autonomous Control Toby Lightheart, Steven Grainger and Tien-Fu Lu School of Mechanical Engineering University of Adelaide, North Terrace, Adelaide, SA 5005 Australia toby.lightheart@adelaide.edu.au Abstract Bertini Jr., 2007]. However, a common drawback of many CoNN techniques is a propensity towards excessive This paper presents a method that draws upon growth and over-fitting. This can be caused by processes reinforcement learning to perform autonomous designed to minimise error during learning that can be learning through the automatic construction of a particularly problematic when learning data is noisy, such spiking artificial neural network. Constructive as is common in real-world applications. neural networks have been applied previously to Standard learning algorithms for ANNs and CoNNs are state and action-value function approximation described as either supervised (training data has inputs but have encountered problems of excessive and desired outputs) or unsupervised (training data only growth of the network, difficulty generalising has inputs) [Haykin, 1994]. For autonomous control