Aprendizado por reforço aplicado ao problema de alocação de berços graneleiros com controle de estoque

Abstract

International maritime transport has established itself as the cornerstone of the global economy, accounting for approximately 80% of the volume and more than 70% of the total value of world trade in goods. Given this magnitude, optimizing port operations efficiency has become strategic for supply chain resilience. However, the recent period (2020–2025) was marked by a succession of disruptive events, including geopolitical crises along critical routes and climatic instability, which posed unprecedented challenges to port logistics. In this context, the Berth Allocation Problem (BAP), although well-established in the literature, demands new approaches that offer greater robustness to uncertainty. Classical optimization methodologies often exhibit scalability limitations and make simplifying assumptions that neglect the high dimensionality and operational volatility of real terminals. This thesis proposes a formulation of the BAP for dry bulk cargo, integrated with inventory control, and modeled under the Deep Reinforcement Learning (DRL) paradigm. The investigation initially employs the DQN (Deep Q-Network) algorithm. Subsequently, it evolves toward an architecture incorporating LSTM (Long Short-Term Memory) cells to capture the temporal dependencies inherent to complex sequential decision-making problems. The experimental results demonstrate that the proposed approach not only generates high-quality solutions in dynamic scenarios but also effectively mitigates inventory failures, surpassing the limitations of traditional heuristics.