Reinforcement learning with continuous vector output

Abstract

A new reinforcement learning algorithm with continuous vector output (CVRL) is proposed for a continuous process with multiple-input and multiple-output. CVRL is a generic hierarchically structured framework. The lower layer is composed with several groups of action units and continuous vector output can be produced based on action combination. The higher layer is a Q-learning unit defined on the space of combined action, its responsibility is the selection of properly combined actions. A detailed implementation of the CVRL is given, and the simulation on a mobile robot navigation problem demonstrates its effectiveness.