Agent transfer learning for cognitive resource management on multi-hop backhaul networks

In this paper we have introduced a transfer learning paradigm for radio resource management applied to a multi-hop backhaul network, which enables the agents (base stations) to share learning knowledge to improve a traditional reinforcement learning algorithm and the system Quality of Service (QoS). We break down transfer learning into three issues with algorithms developed for each: a source agent selection scheme that groups the neighbour links to exchange learning information; a target agent training scheme that enhances the learner's knowledge base by training functions; an information exchange control scheme that controls the cooperation overhead. It is validated that by transferring the weight table conversely from neighbour source agents, the target agents can make more accurate resource allocation decision, with up to 50% reduction in retransmissions. Moreover, the transfer process can be terminated once the target agent has mature knowledge from source agents, which reduces up to 95% of control information overhead with efficient QoS achieved.