Learning-based ship design optimization approach

With the development of computer applications in ship design, optimization, as a powerful approach, has been widely used in the design and analysis process. However, the running time, which often varies from several weeks to months in the current computing environment, has been a bottleneck problem for optimization applications, particularly in the structural design of ships. To speed up the optimization process and adjust the complex design environment, ship designers usually rely on their personal experience to assist the design work. However, traditional experience, which largely depends on the designer’s personal skills, often makes the design quality very sensitive to the experience and decreases the robustness of the final design. This paper proposes a new machine-learning-based ship design optimization approach, which uses machine learning as an effective tool to give direction to optimization and improves the adaptability of optimization to the dynamic design environment. The natural human learning process is introduced into the optimization procedure to improve the efficiency of the algorithm. Q-learning, as an approach of reinforcement learning, is utilized to realize the learning function in the optimization process. The multi-objective particle swarm optimization method, multi-agent system, and CAE software are used to build an integrated optimization system. A bulk carrier structural design optimization was performed as a case study to evaluate the suitability of this method for real-world application.

► This paper presents a new machine-learning-based optimal ship design approach, which introduces the human learning process into the practical ship design and analysis to improve the efficiency of optimization.
► QQ-learning is employed to solve the sensory memory and partial short time memory problem via a multi-agent system.
► This method can guide the direction of optimization via experience.
► It greatly improves the performance of optimization in terms of accuracy while and it reduces the computing time by 25%.