A Dueling Double Deep Q-Network-Assisted NSGA-II Framework for Multi-objective Ship Subdivision Optimization

Ship subdivision optimization is challenging because it involves a high-dimensional mixed-integer design space and requires repeated strength evaluations. To address these challenges in preliminary ship design, this study develops a two-level parametric modeling strategy based on an Inner Shell Control Entity (ISCE). This strategy provides a unified and flexible description of inner-shell geometry and transverse bulkhead arrangements, enabling the automatic generation of three-dimensional compartment configurations suitable for strength, hydrostatic, and stability analyses. Building on this parametric representation, a multi-objective optimization framework combining NSGA-II with a Dueling Double Deep Q-Network (D3QN) is proposed to improve the search efficiency of ship subdivision design. The optimization model aims to maximize total cargo volume while minimizing still-water bending moment under constraints of geometric feasibility, compartment capacity allocation, hydrostatics and stability, and longitudinal strength. The evolutionary process is formulated as a Markov decision process. A D3QN agent adaptively selects crossover and mutation probability combinations according to evolutionary state indicators, thereby achieving an effective balance between exploration and exploitation. Case studies on a 55,000 DWT product oil tanker and a 50,000 DWT bulk carrier demonstrate that the proposed framework attains Pareto-optimal solutions comparable to those obtained by conventional NSGA-II and multi-objective particle swarm optimization (MOPSO), while significantly accelerating convergence. Specifically, the number of generations required to reach 99% of the maximum hypervolume is reduced by 83.9% and 77.0%, respectively, compared with standard NSGA-II. Furthermore, the D3QN agent trained on the tanker case is directly transferred to the bulk-carrier case without retraining, demonstrating cross-ship-type generalization capability. The optimized subdivision schemes provide cargo-volume adjustment ranges of 5.7% and 8.0% without altering the principal dimensions, confirming the practical applicability of the proposed method in supporting engineering trade-offs between payload capacity and longitudinal strength. • A two-level parametric subdivision model based on Inner Shell Control Entity is proposed. • D3QN-assisted NSGA-II enables adaptive crossover-mutation parameter selection. • Convergence efficiency improves by 83.9% and 77.0% for tanker and bulk carrier cases. • Subdivision optimization provides over 6% cargo-volume adjustability. • The framework demonstrates ship-type generality across different vessel designs.