논문 상세보기
Reinforcement Learning-Based Multi-Objective Arctic Route Optimization Considering Sea Ice Risk, Time, and Cost KCI 등재
해빙 위험도와 시간ㆍ비용을 고려한 강화학습 기반 북극항로 다목적 최적화
- 언어KOR
- URLhttps://db.koreascholar.com/Article/Detail/452101
pp.131-141
The Northern Sea Route (NSR), increasingly navigable as Arctic sea ice retreats, can substantially shorten shipping distance, but route planning is difficult because navigability and risk at a location depend on a ship's arrival time. In practice it is rarely a single-route problem: forecasts are uncertain and frequently updated, and operators must weigh many departure times and time-cost trade-offs, so a planner is queried repeatedly. Graph-search and metaheuristic methods recompute a full solution per query and become the bottleneck, while prior reinforcement learning studies on Arctic routes assume static ice and ignore arrival-time-dependent risk. This study proposes a framework in which ST-A* (Space-Time A*) generates demonstration trajectories for different time-cost weights, and DQfD (Deep Q-Learning from Demonstrations) uses them to learn weight-specific policies that produce routes for new ice conditions by inference, without the full space-time re-search. On 51 unseen test cases from 2023-2025, the learned policies reached the destination in every case with lower objective values than a baseline trained without demonstrations. Relative to ST-A*, they produced routes about 700 times faster at a mean objective 1.57 times higher, a bounded loss in per-route optimality that makes large numbers of route evaluations tractable.
2. 문헌 연구
3. 제안 방법
3.1 POLARIS 기반 해빙 위험도 평가
3.2 목적함수 및 제약조건
3.3 비용 모델
3.4 강화학습 환경 및 리워드 설계
3.5 Space-Time A* 기반 데모 경로 생성
3.6 DQfD 기반 정책 학습
4. 실험 설계
4.1 데이터 구성
4.2 비교 대상 및 평가 지표
5. 실험 결과
5.1 다목적 경로의 상충관계
5.2 Space-Time A* 대비 성능
5.3 강화학습 baseline 비교
6. 함의
6.1 연산 효율과 경로 품질 상충관계 및 활용 방안
6.2 데모 활용의 도달성 기여
7. 결론 및 향후 연구
Acknowledgement
References
