Q-learning-based data replication for highly dynamic distributed hash tables

This paper focuses on data replication in structured peer-to-peer systems over highly dynamic networks. A Q-learning-based replication approach is proposed. Data availability is periodically computed using the Q-learning function. The reward/penalty property of this function attenuates the impact of the network dynamism on the replication overhead. Hence, the departure of a node does not necessarily lead to the addition of a replica in the network. The replication process is triggered according to the overall data availability. Simulation results proved that the proposed approach ensures data availability in dynamic environments with minimum data transfer costs.