Adversarial Inverse Reinforcement Learning for Energy-Efficient Marine Vessel Operation

Abstract

Abstract The maritime industry serves as a backbone of global trade, transporting nearly 80% of goods worldwide, making it indispensable for the global economy (Statista,2024; UNCTAD, 2024). Yet, according to UNCTAD, this sector remains a notable contributor to greenhouse gas (GHG) emissions, accounting for 2-3% of total emissionsa figure projected to rise dramatically by 2050 under business-as-usual scenarios. Addressing this escalating environmental impact requires innovative strategies for optimizing vessel operations and minimizing fuel consumption, which are crucial for sustainable maritime practices. This thesis focuses exclusively on the application of adversarial inverse reinforcement learning (AIRL) to tackle these challenges. AIRL leverages expert demonstrations from seasoned captains who have successfully navigated complex and diverse maritime conditions. By analyzing these expert behaviors, AIRL derives comprehensive reward function that encapsulate optimal operational strategies, paving the way for the development of a control policy. The strength of AIRL lies in its ability to capture the nuances of expert judgment, enabling marine vessels to replicate the energy-efficient decisions made by experienced captains. Unlike traditional control systems, which often struggle with the nonlinear dynamics and diverse challenges of maritime environments, AIRL offers a robust data-driven solution tailored to the specific operational needs of the vessels. By integrating AIRL, this research aims to enhance energy efficiency, reduce fuel consumption, and lower carbon emissions. In doing so, it contributes not only to the operational goals of modern marine vessels but also to the broader environmental objectives of the global community. This work aligns with the maritime industry’s transition toward greener practices and supports the adoption of sustainable solutions.