Continuous Q-Learning Resource Allocation Network

摘要

This paper presents a learning architecture based on Q-Learning for learning control problems with multi-dimensional continuous state and action spaces. CQRAN (Continuous Q-Learning Resource Allocation Netowork) consists of two integrated function approximators for state and action space representation. Using dynamic resource allocation allows the construction of task-specific Radial Basis Function networks. The architecture supports a direct determination of the executing action. These two aspects of the system allow the efficient application on robot tasks. CQRAN is tested with common benchmark problems in the robotic field.