상세 보기
A unified learning framework for the block relocation problem: A new performance baseline
- Shin, Woo-Jin;
- Jung, Ji-Kwang;
- Cho, Sang-Hyun;
- Choi, Inguk;
- Tanaka, Shunji;
- 외 1명
SCOPUS
0초록
This study addresses the block relocation problem (BRP), a key challenge in container yard operations. In the BRP, containers are stacked vertically, and retrieving a target container may require relocating those above it. The objective is to retrieve all containers in a predefined order while minimizing the total number of relocations. Recent reinforcement learning (RL) approaches have been explored for the BRP, but they face fundamental limitations. In particular, the unrestricted variant allows any topmost container to be relocated, which may lead to cyclic relocations, non-termination, and arbitrarily long decision trajectories. These issues can produce unreliable learning signals and substantial training-time and memory overhead. To address these challenges, we propose L2BR, a unified learning framework for both the restricted and unrestricted BRP that sequentially combines imitation learning (IL) and self-improvement learning (SIL). IL mitigates cycle-related issues by training an initial policy from pre-generated (near-)optimal solutions, while SIL enables scalable improvement by imitating the best complete solutions found by the policy itself. Experiments on well-known benchmark sets show that L2BR consistently finds optimal solutions for many instances and establishes new best-known results within only a few seconds. L2BR substantially outperforms existing RL-based methods and also surpasses state-of-the-art optimization-based approaches, particularly for the unrestricted variant and large-scale instances.2 © 2026 Elsevier B.V.
키워드
- 제목
- A unified learning framework for the block relocation problem: A new performance baseline
- 저자
- Shin, Woo-Jin; Jung, Ji-Kwang; Cho, Sang-Hyun; Choi, Inguk; Tanaka, Shunji; Kim, Hyun-Jung
- 발행일
- 2026
- 유형
- Article in press
- 언어
- ENG
- 출판사
- Elsevier B.V.
- 발행국가
- 네덜란드
- ISSN
- E 1872-6860
P 0377-2217