PROXIMAL POLICY OPTIMISATION VERSUS ANT COLONY OPTIMISATION FOR THE THREE-DIMENSIONAL BIN PACKING PROBLEM: A COMPARATIVE STUDY
PROXIMAL POLICY OPTIMISATION VERSUS ANT COLONY OPTIMISATION FOR THE THREE-DIMENSIONAL BIN PACKING PROBLEM: A COMPARATIVE STUDY

Summary/Abstract: This paper compares a Proximal Policy Optimisation (PPO) deep reinforcement-learning agent with an Ant Colony Optimisation (ACO) solver on the offline, heterogeneous-bin three-dimensional bin packing problem (3D-BPP). Both algorithms were evaluated on fifty synthetic instances using a unified composite scoring function covering placement ratio, volume utilisation, bin-count penalty and mean per-bin waste. PPO achieves a higher mean composite score (0.346 vs. 0.283), wins on 38 of 50 instances with an average winning margin of 0.101, and resolves each instance in under 60 seconds on a commodity CPU. ACO exhibits greater score variance and resolves instances in up to 1,706 seconds, but its training-free character makes it relevant when the instance distribution changes too rapidly for policy retraining. The PPO training cost of approximately 5.5 hours is recovered after 58 instances compared with ACO at mean inference times. A paired Wilcoxon signed-rank test is identified as the appropriate significance test once per-instance data are made available.

• Details
• Contents