A Q-learning system considering swing-up gains for controlling a parallel double inverted pendulum

Abstract In this paper, a new reinforcement learning technique is proposed to obtain autonomously the swing-up gains and switching areas of local controllers for controlling serial double inverted pendulum. The proposed method integrates learning method for the swing-up gains and switching control using conventional local controllers. Local controllers are used a priori knowledge of the plant so that the admissible control performance can be obtained quickly. By using the proposed method, appropriate swing-up gains and switching area for controllers can be obtained for swing-up and stabilization. Then, the control objective can be achieved by smallest swing-counts of pendulums. Moreover lookup tables used in the learning method are constructed by CMAC that is table-lookup method with generalization capability, so that the number of learning iterations in early stages of learning can be reduced.