作者： ANDREW BARTO
DOI:
关键词:
摘要: Considerable progress was made in developing artificial neural network methods for solving stochastic sequential decision problems. The research focused on reinforcement learning methods based on approximating dynamic programming (DP). They used problems in the domains of robot fine motion control, navigation, and steering control in order to develop and test learning algorithms and architectures. Although most of these problems were simulated, they also began to apply DP-based learning algorithms to actual robot control problems with considerable success. Progress was made on reinforcement learning methods using continuous actions, modular network architectures, and architectures using abstract actions. Theoretical progress was made in relating DP-based reinforcement learning algorithms to more conventional methods for solving stochastic sequential decision problems. As a result of this …