多Agent MDPs中并行Rollout学习算法

文章在rollout算法基础上研究了在多Agent MDPs的学习问题.利用神经元动态规划逼近方法来降低其空间复杂度,从而减少算法"维数灾".由于Rollout算法具有很强的内在并行性,文中还分析了并行求解方法.通过多级仓库库存控制的仿真试验,验证了Rollout算法在多Agent学习中的有效性.

Parallel rollout algorithms for multi-agent MDPs

The paper researches Rollout algorithms （RA） for multi-Agent Markov decision processes （MDPs） in the framework of performance potentials theory. Neuro-dynamic programming （NDP） is used to reduce ＂curse of dimensionality＂ of algorithms, Since to rolout algorithms has a very strong intrinsic parallelism,the parallelization method of RA is employed to reduce the time of running algorithms. Finally,an example of multi-level inventory control by using RA under the supply chain environment is provided. The result shows that rollout algorithms are confirmed to be valid in multi-Agent learning.