Novel Scheme For Congestion Control In Cellular Networks Using Deep Reinforcement Learning And Markov Decision Process Models

This research deals with the general issue of quality of service (QoS) provisioning and resource utilization in telecommunication networks. The issue requires that mobile network income be optimized while simultaneously satisfying QoS constraints that prevent getting into specific states and utilization of specific actions. However, supporting QoS requirements of different traffic types is more complicated due to the need to minimize two performance indicators - the probability of discarding a handover call and the probability of hindering a new call. Several approaches proposed recently try to provide efficient model-based solution to the problem by formulating it as an average reward neuro-dynamic programming (NDP) optimization problem together with decomposition function, but this is limited by Bellman’s curse of dimensionality. In this paper, we proposed a novel hybrid optimization scheme to address the problem using Deep Reinforcement Learning (DRL), Markov Decision Process (MDP) and adaptive joint call admission control (AJCAC) respectively. In the proposed scheme, two classes of arrival traffic at the base station (BS) are considered; voice (real-time) and data (non-real-time) calls. Furthermore, traffic is classified as new and handoff according to the type of request. The scheme introduces an adaptive threshold value, which dynamically adjusts the network resources under high traffic intensity. In addition, the scheme introduces a learning agent whose state is described by an MDP. The MATLAB version 2010 software, OMNET++ simulator and SPSS will be used for data, numerical, algorithm simulation and modeling. Data analysis and simulation results will be carried out for performance evaluation of the proposed DQL-AJCAC scheme against existing models.