LEARNING-GUIDED SEARCH FOR RELOCATION PROBLEM IN MULTI-DEEP AVS/RS SYSTEMS

Abstract

Efficient relocation handling is essential in multi-deep automated vehicle storage and retrieval systems (AVS/RS), where requested stock keeping units are frequently blocked by other items stored in front of them. As warehouse depth and occupancy increase, the resulting relocation problem becomes highly combinatorial and difficult to solve with exact optimization methods. This paper proposes a learning-guided search framework for the restricted relocation problem in multi-deep AVS/RS systems. The approach combines Reverse Renumbering Relocation Construction (R3C), supervised value learning, and beam search. First, R3C generates synthetic warehouse states with known relocation counts by construction, enabling large-scale supervised learning without repeated exact optimization. Second, a DeepSets-style ValueNet is trained to estimate the remaining number of relocations from a given system state. Third, the learned value model is embedded into a beam search procedure that prioritizes promising relocation sequences. The experimental analysis focuses on a medium-sized 5x40 warehouse and shows that enlarging the synthetic training dataset from 45,000 to 135,000 instances improves validation performance from RMSE 0.712 and MAE 0.535 to RMSE 0.589 and MAE 0.475. This improvement also reduces the average search gap to the R3C reference from 14.78 to 7.62 relocations under the same beam-search setting. The results further show that the proposed framework performs very well at lower occupancy levels, while dense warehouse states remain significantly more challenging. The study demonstrates that supervised learning can provide effective search guidance for relocation decision-making in complex AVS/RS environments.