지원사업
학술연구/단체지원/교육 등 연구자 활동을 지속하도록 DBpia가 지원하고 있어요.
커뮤니티
연구자들이 자신의 연구와 전문성을 널리 알리고, 새로운 협력의 기회를 만들 수 있는 네트워킹 공간이에요.
논문 기본 정보
- 자료유형
- 학술대회자료
- 저자정보
- 발행연도
- 2023.11
- 수록면
- 735 - 741 (7page)
이용수
초록· 키워드
오류제보하기
The growing demand for autonomous vehicles emphasizes the need for safe emergency operations. Minimal risk maneuver (MRM) is an answer to such demand since it can warrant safe emergency operations of autonomous vehicles, safeguarding against potential hazards and ensuring passenger safety. Existing safe emergency operations algorithms often rely on heuristic or deterministic models. Reinforcement learning offers the potential to improve MRM in an iterative manner, based on environmental feedback. In this paper, we examine the possibility of MRM algorithms empowered by reinforcement learning. Our experimental evaluation is conducted in three stages. In the first stage, the parameter space for MRM is defined where the parameter space is all possible combinations of relevant variables characterizing the environment and conditions of an automated driving system (ADS). Such variables are environmental conditions, traffic conditions,vehicle-specific parameters, sensor data, regulatory and legal constraints, and road infrastructures. In the second stage, we select three representative reinforcement learning (RL) algorithms, Trust Region Policy Optimization (TRPO), Proximal Policy Optimization (PPO), and Policy Gradient (PG), to train our MRM model. We then generate three distinct MRM agents that are trained by the three algorithms, respectively. In the final stage, we run these MRM agents with test scenarios in a simulation environment and measure their performance with a specific emphasis on their ability to handle emergency situations and move safely onto Minimal Risk Condition (MRC). Specifically, MRM safety metrics, including reinforcement learning performance, total travel time, and MRM success rates, are quantitatively gathered to assess the three agents. The experiment results show the MRM agent trained by TRPO outperforms the other two trained by other algorithms.
목차
Abstract
1. INTRODUCTION
2. BACKGROUND
3. PROBLEM OVERVIEW
4. PARAMETER SPACE
5. TRAINING PHASE OF RL
6. EVALUATION ON SIMULATION
7. RELATED WORK
8. CONCLUSION
References
참고문헌 (0)
참고문헌 신청이 논문의 저자 정보
이 논문과 함께 이용한 논문
최근 본 자료
댓글(0)
0