TY - GEN
T1 - CNN-based Apprenticeship Learning for Inverse Reinforcement Learning
AU - Kim, Ye Rin
AU - Choi, Hyun Duck
N1 - Publisher Copyright:
© 2024 ICROS.
PY - 2024
Y1 - 2024
N2 - In this paper, we propose a CNN-based inverse reinforcement learning method that optimizes a reward function modeled by a linear combination. The proposed method efficiently extracts features from expert demonstrations using a CNN-based network and effectively estimates the reward function with a few iterations. The proposed method is called CNN-based apprenticeship learning for inverse reinforcement learning. The policy estimated by this method guarantees performance similar to or better than that of expert behavior. Through the Super Mario simulation, we demonstrate that the proposed CNN-based apprenticeship learning outperforms traditional imitation learning and reinforcement learning methods.
AB - In this paper, we propose a CNN-based inverse reinforcement learning method that optimizes a reward function modeled by a linear combination. The proposed method efficiently extracts features from expert demonstrations using a CNN-based network and effectively estimates the reward function with a few iterations. The proposed method is called CNN-based apprenticeship learning for inverse reinforcement learning. The policy estimated by this method guarantees performance similar to or better than that of expert behavior. Through the Super Mario simulation, we demonstrate that the proposed CNN-based apprenticeship learning outperforms traditional imitation learning and reinforcement learning methods.
KW - Apprenticeship Learning via Inverse Reinforcement Learning (ALIRL)
KW - Convolutional Neural Network (CNN)
KW - Deep Q-Network (DQN)
UR - http://www.scopus.com/inward/record.url?scp=85214351221&partnerID=8YFLogxK
U2 - 10.23919/ICCAS63016.2024.10773173
DO - 10.23919/ICCAS63016.2024.10773173
M3 - Conference contribution
AN - SCOPUS:85214351221
T3 - International Conference on Control, Automation and Systems
SP - 73
EP - 78
BT - 2024 24th International Conference on Control, Automation and Systems, ICCAS 2024
PB - IEEE Computer Society
T2 - 24th International Conference on Control, Automation and Systems, ICCAS 2024
Y2 - 29 October 2024 through 1 November 2024
ER -
