| Title |
Robot Arm Control Technique using Deep Reinforcement Learning based on Dueling and Bottleneck Structure |
| Authors |
김성준(Seong Joon Kim) ; 김병욱(Byung Wook Kim) |
| DOI |
https://doi.org/10.5370/KIEE.2021.70.12.1906 |
| Keywords |
Deep reinforcement learning; Q-learning; Dueling network; Bottleneck layer; Robot arm control |
| Abstract |
In this paper, we propose a deep reinforcement learning network using dueling and bottleneck structure to improve the task completion rate and computational efficiency of a robot arm control. The bottleneck structure applied to the neural network reduces the number of parameters and the amount of computation by adding 1 * 1 convolution and 3 * 3 convolution to the output layer.
In addition, by applying the dueling structure to the neural network and dividing the function into an advantage function and a value function, it prevents the bad action selection that can occur in existing learning and reduces the variance of value, thereby improving learning stability and estimation of the agent's state. As a result of the experiment using the V-REP simulator, the proposed method improves the work completion rate and work efficiency by approximately 10.3% and 1.6%, respectively, while reducing the number of parameters by 35.9% compared to the existing VPG network. |