Reinforcement Learning for Active Queue Management in Mobile All-IP Networks

In future all-IP based wireless networks, like the envisaged in the long term evolution (LTE) architectures for future systems, network providers will have to deal with large traffic volumes with different QoS requirements. In order to increase exploitation of network resources wisely, intelligent adaptive solutions for class based traffic regulation are needed. In particular, active queue management (AQM) is regarded as one of these solutions to provide low queuing delay and high throughput to flows by smart packet discarding. In this paper, we propose a novel AQM solution for future all-IP networks based on a reinforcement learning scheme that allows controlling both the queuing delay and the packet loss of the different service classes. The proposed approach is evaluated through simulations and compared against other algorithms used in the literature, like the random early detection (RED) and the drop from tail (DFT), confirming the benefits of the proposed algorithm.