Abstract: Based on the Markov property of the growth steps of shell/plate stiffeners, a reinforcement learning driving strategy of the growth design of shell/plate stiffeners was proposed. Aiming at minimizing the overall strain energy of the structures, Markov decision process was used to model the growth processes of the stiffeners. By introducing a multi-agent system to share the states and the rewards of the stiffeners growth processes, and memorizing specific actions, the learning complexity was reduced. Meanwhile, the convergence of the reward value of the stiffeners growth processes was realized. Therefore, the growth design strategy of shell/plate stiffeners was achieved. Finally, a numerical example was given and the results of the smoothed stiffeners layout were compared with those of the classical algorithm, which verifies the validity of the growth design of stiffeners for shell / plate structures based on MADQN interaction.

Key words:  , stiffener for shell/plate structure, growth pattern, multi-agent deep Q network(MADQN), layout design, reinforcement learning

摘要： 基于板壳加强筋生长步序列的马尔可夫性质，提出了板壳加强筋生长式设计的强化学习驱动策略。以结构整体应变能最小化为目标，运用马尔可夫决策过程对板壳加强筋的生长过程进行建模。通过引入多智能体系统，共享加强筋生长式过程的状态奖励并记忆特定动作，降低学习复杂度，实现了加强筋生长式过程奖励值的波动收敛，达成板壳加强筋生长式设计策略。最后给出算例并将平滑处理后的加强筋布局与经典算法的设计结果进行对比，验证了基于多智能体深度Q网络交互的板壳加强筋生长式设计的有效性。

关键词: 板壳加强筋, 生长式, 多智能体深度Q网络, 布局设计, 强化学习