Abstract: This paper proposes a real-time scheduling model based on contract net protocol structure employing reinforcement learning agents. To this end, an R-learning procedure is elaborated and embedded in machine agents’ decision process, enabling them to treat bid-invitations in more complicated way than in a simple contract net protocol environment. Efficiency of the proposed method is verified through experiments in a simulated real-time scheduling environment. Furthermore, the performance of mixed machine groups which comprises both reinforcement learning agents and non-reinforcement-learning agents shows that there is spontaneous implicit teamwork occurring between reinforcement learning agents, and this teamwork guarantees high quality output of the scheduling model.

Key words: R-learning, contract net protocol, multi-agent cooperation, real-time schedule

摘要： 提出一种融入合同网运行机制的R学习方法，以此方法为核心构造Agent形成具有学习能力的实时调度模型。模型以最小化作业累计平均流动比为主要目标，同时借助对强化学习报酬的设计减小机器负载的不均衡性，实现对调度过程的双重优化；构造实时调度实例投入测试的结果证明了模型的绩效。另外，一个包含强化学习Agent与无学习Agent的混合机器环境被构建并测试其性能，测试结果表明：在Agent之间借助强化学习过程形成了某种隐性的合作，正是这种合作保证了高质量实时调度方案的输出。

关键词: R学习, 合同网, 多Agent合作, 实时调度