Development of deep reinforcement learning for inverted pendulum

This paper presents a modification of the deep Q-network (DQN) in deep reinforcement learning to control the angle of the inverted pendulum (IP). The original DQN method often uses two actions related to two force states like constant negative and positive force values which apply to the cart of IP to maintain the angle between the pendulum and the Y-axis. Due to the changing of too much value of force, the IP may make some oscillation which makes the performance system could be declined. Thus, a modified DQN algorithm is developed based on neural network structure to make a range of force selections for IP to improve the performance of IP. To prove our algorithm, the OpenAI/Gym and Keras libraries are used to develop DQN. All results showed that our proposed controller has higher performance than the original DQN and could be applied to a nonlinear system.