Reinforcement learning in production logistics management at shipbuilding enterprises: comparative analysis of results and development prospects

Abstract

The paper presents a systematic analysis of the current state of research on the application of reinforcement learning methods to optimize production logistics and supply chain management in the shipbuilding industry. The specific features of shipbuilding as engineer-to-order production are described in detail; in this type of production, each product is designed and manufactured from scratch in accordance with unique customer requirements, which distinguishes it from serial production. It is shown that this feature places increased demands on the adaptability of planning systems, which must operate under conditions of uncertainty and frequent changes. A comparative analysis is conducted of three main classes of optimization methods: classical heuristic dispatching rules, metaheuristic approaches, including genetic algorithms and swarm intelligence algorithms, and various types of deep reinforcement learning algorithms. The methods are compared according to important criteria of solution quality, computational efficiency, robustness to disturbances, and the ability to continue learning in real time. It is noted that reinforcement learning provides greater adaptability and faster decision making than traditional methods, especially in a dynamically changing production environment. Special attention is paid to digital twins as an essential environment for the safe training of agents in accordance with the requirements of the international standard; it is shown that only such an environment makes it possible to conduct repeated experiments without interrupting actual production. Examples of the implementation of reinforcement learning elements at shipyards and in related industries, including pilot projects in South Korea, Spain, and Norway, are systematized. Based on the identified gaps, promising research directions are formulated: the development of multi-agent systems for coordinating multiple decisions, the use of transfer learning for accelerated adaptation to new order types, and integration with generative artificial intelligence for synthesizing alternative planning scenarios.