Multi-agent Learning for Resource Allocationn Dense Heterogeneous 5G Network

Abstract

The concept of small cells challenges many timehonored assumptions about the structure of cellular networks. In this approach the large number of low-power devices is deployed to increase the spatial frequency reuse of the selected area. The efficient resource management and interference coordination schemes became an important requirement for the successful adoption of heterogeneous networks. In this paper we propose a distributed multi-agent strategy, where small cells locally control resource usage, such that the overall system capacity is maximized. The main goal resides in providing each cell with the ability to make its decision autonomously while dynamically taking into account the resource occupation of the surrounding cells. We study an area of coexistence with non-cooperative macroenvironment and propose a mechanism to increase the efficiency of learning with a smart safe-shift phase. We illustrate the application of this distributed learning strategy for the subband allocation and propose several mechanisms to improve the convergence speed in the absence of communication. The performances of the proposed method are evaluated in the case of Long Term Evolution (LTE) setup and compared to a number of resource allocation schemes. We validate the algorithm with system level simulations and show that it achieves considerable improvement in system performance for heterogeneous deployment with non-cooperating agents, without compromising the efficiency of the system.