자율주행 엣지의 CaSSLe 기반 오프라인 강화학습 모델 설계

초록

There is growing interest in applying Reinforcement Learning (RL) to autonomous driving edge devices, which require both safety and real-time performance. We introduce a continual learning algorithm, CaSSLe (Continual Self-Supervised Learning), integrated with Double DQN-based reinforcement learning to enhance adaptability to new situations. Our proposed algorithm, AP-CaSSLe (Adaptive Progressive-Continual Self-Supervised Learning), dynamically adjusts the strength of input data augmentation. The AP-CaSSLe algorithm demonstrated a 58.9% improvement in accuracy compared to the Double DQN without continual learning, highlighting its potential for enhancing performance in autonomous driving edge environments.

키워드