Abstract: In order to reduce the overhead caused by real-time network monitoring, a policy for fault monitoring of a hierarchical network with a multi-manager is proposed. This fault monitoring policy is based on the model of multi-agent Markov Decision Processes and makes use of the reinforcement learning mechanism. Simulation results show that the proposed scheme based on MMDP makes the network management more efficient by accurately detecting the faulty node and significantly reducing the management overhead.

Key words: network fault management, policy management, MMDP, reinforcement learning