| Title |
Cart-Pole System Control Using Memory Transformer Q-Learning |
| Authors |
한병찬(Byeong-Chan Han) ; 강민제(Min-Jae Kang) ; 송성호(Seong-Ho Song) ; 김호찬(Ho-Chan Kim) |
| DOI |
https://doi.org/10.5370/KIEE.2024.73.12.2371 |
| Keywords |
Deep reinforcement learning; DQN; DDQN; Dueling DDQN; Transformer |
| Abstract |
This paper proposes a memory transformer Q-learning network(MTQN) algorithm to improve existing deep reinforcement learning algorithms. MTQN is configured by combining transformers with existing deep reinforcement learning models to model sequence systems more efficiently, and the gating mechanism of LSTM is additionally used for using the transformer. The proposed algorithm is compared and analyzed with DQN, a representative deep reinforcement learning algorithm, and its modified algorithms, targeting cart-pole system, a representative reinforcement learning benchmark environment. The simulation extracts and compares the evaluation score, cart position, and pole angle of cart-pole system, and shows that the proposed algorithm learns the fastest and most stably. |