Optimizing one-dimensional bin packing for conveyor belt logistics via pointer-network-based A2C and heuristics

The one-dimensional Bin Packing Problem (1D-BPP) poses significant challenges in logistics, particularly for optimizing item packing sequences in warehouse conveyor belt systems. This paper proposes a Deep Reinforcement Learning (DRL) approach using Advantage Actor-Critic (A2C) integrated with Pointer Networks to learn optimized item sequencing policies for 1D-BPP. Our method generates item permutations that are subsequently processed by First-Fit , Next-Fit , or Best-Fit heuristics, effectively bridging the gap between offline sequence optimization and online packing constraints. The Actor Network, incorporating a Pointer Network with an LSTM-based encoder-decoder and attention mechanism, learns a stochastic policy to produce permutations that improve bin occupancy. The Critic Network estimates the state-value function to reduce variance during policy gradient updates. Experimental results on small (5–10 items), medium (20–30 items), and large (100–200 items) problem instances demonstrate that DRL-enhanced First-Fit and Best-Fit achieve statistically significant improvements in average bin occupancy over standalone heuristics and First-Fit-Decreasing baselines ( \(p < 0.001\) ), with effect sizes ranging from moderate to very large. The proposed framework offers a computationally efficient approach to sequence optimization for bin packing in dynamic warehouse environments.