考虑劣化维护的单机调度深度强化学习模型和算法

陈勇,杜习之,姜一炜,易文超,裴植,纪祖臻

Deep reinforcement learning models and algorithms for single-machine scheduling considering deteriorated maintenance

Yong CHEN,Xizhi DU,Yiwei JIANG,Wenchao YI,Zhi PEI,Zuzhen JI

表 6 DRL算法训练性能比较

Tab.6 Training performance comparison of DRL algorithms

规模 算法 时间步/106 平均FPS 训练时长/h 10 A2C 1 1498.4 0.18 DQN 1 1660.8 0.27 PPO 1 1141.6 0.24 20 A2C 2 1473.7 0.37 DQN 2 1440.6 0.54 PPO 2 1109.6 0.50 30 A2C 2 1447.3 0.38 DQN 2 1393.9 0.54 PPO 2 1093.3 0.51 50 A2C 4 1400.9 0.79 DQN 4 1283.4 1.11 PPO 4 1065.4 1.05 80 A2C 8 1360.6 1.63 DQN 8 1154.3 2.27 PPO 8 1038.6 2.15 100 A2C 8 1350.4 1.66 DQN 8 1121.4 2.31 PPO 8 1018.6 2.19 150 A2C 16 1276.1 3.51 DQN 16 1051.9 4.68 PPO 16 1007.2 4.33