Reinforcement learning based base station cooperation scheme in mobile networks

Abstract

Recently, the development of smart devices causes a lot of traffic. To deal with the increment of traffic, high data rate is required in mobile networks. There are many schemes to achieve the high data rate. In this thesis, we adopt base station (BS) cooperative communication scheme. In BS cooperative communication, lots of problems occured. Among these problems, we are interested in making combinations of BSs that are participated in the same cooperation. This combination of the BSs are called cluster. There are two popular schemes in BS clustering issue. One is static clustering, and the other is dynamic clustering. In static clustering, the cluster set is fixed. Their objective is usually maximizing SINR or minimizing outage probability. Static clustering has no computation complexity during service because the cluster set does not change. However, the performance gain is not stable. It depends on the user distribution. When the mobile stations (MSs) are concentrated at the edge of cluster, there is no performance gain due to clustering. On the other hand, dynamic clustering uses instantaneous channel state information. It calculates joint SINR for every BS combination and then finds appropriate BS combination which achieves maximum joint SINR. The performance gain of dynamic clustering is greater than that of static clustering. However, computational complexity of dynamic clustering is very high. It is hard to implement dynamic clustering in real environment. Between the concept of static and dynamic clustering, semi-dynamic clustering is proposed. In semi-dynamic clustering, the system has a number of static clustering sets. The system dynamically chooses one static clustering set in a time slot. Semi-dynamic clustering can get the advantage from the dynamical property of dynamic clustering while reducing the computational complexity. In this thesis, we propose a novel algorithm for managing the semi-dynamic cluster. We apply a reinforcement learning algo...