Research On Quick Dynamic Interference Coordination Scheme In Ultra-Dense Networks

With the popularity of intelligent mobile terminals and the Internet of things,wireless mobile communication systems become increasingly ultra-dense.It is expected that in the next generation mobile communication systems,the number of terminals all over the world will exceed 10 billions,and the number of connections per square kilometer will exceed 1 million.Due to the scarcity of spectrum,for communications involving in large-scale devices,it is necessary to use frequency reuse technique to make adjacent cells reuse the same spectrum.However,frequency reuse causes severe cochannel interference,and interference coordination must be used to decrease interference and improve system spectral efficiency.Dynamic interference coordination has received widespread attention due to its high flexibility.To realize a dynamic interference coordination scheme in a communication system,it is reuqired that its scheduling cycle should be on the level of milliseconds,but traditional schemes take too long computational time in ultra-dense networks.Therefore,this dissertation proposes a quick dynamic interference coordination schem suitable for ultra-dense networks.In scenarios that maximize single-cell performance,such as stadium and convention center,terminals around these sites are relatively sparse,so a central cell may be given the highest priority during optimization.Therefore,a quick interference coordination scheme for such scenarios is theoretically proposes.When the utility function satisfies the conditions of our theorem,the scheme only needs to rank several vectors of channel state parameters to obtain the optimal coordination.In scenarios that maximize the capacity of the whole celluar network,this dissertation combines discrete wavelet transform and deep reinforcement learning to design a generic scheme.Discrete wavelet transform can decrease the dimension of feature parameters and decrease the scheme's computational complexity.Deep reinforcement learning can obtain an optimal coordination through a "self-learning" mechanism.This scheme first compresses the state space and the action space.Then during the training phase,a deep Q network is trained offline through interactions between the agent and the environment.In the application phase,the agent adaptively obtains the optimal action selection strategy based on the trained deep Q network.Finally,an interference coordination result can be obtained through discrete wavelet inverse transform.According to the simulation results,compared with the traditional schemes,the time complexity of this scheme is greatly reduced,and the system performance loss is quite limited,so it has obvious advantages in the scenario with a large number of terminals.