基于GRU和LSTM组合模型的车联网信道分配方法

针对车联网中高通信需求和高移动性造成的车对车链路(Vehicle to Vehicle, V2V)间的信道冲突及网络效用低下的问题，提出了一种基于并联门控循环单元(Gated Recurrent Unit, GRU)和长短期记忆网络(Long Short-Term Memory, LSTM)的组合模型的车联网信道分配算法。算法以降低V2V链路信道碰撞率和空闲率为目标，将信道分配问题建模为分布式深度强化学习问题，使每条V2V链路作为单个智能体，并通过最大化每回合平均奖励的方式进行集中训练、分布式执行。在训练过程中借助GRU训练周期短和LSTM拟合精度高的组合优势去拟合深度双重Q学习中Q函数，使V2V链路能快速地学习优化信道分配策略，合理地复用车对基础设施(Vehicle to Infrastructure, V2I)链路的信道资源，实现网络效用最大化。仿真结果表明，与单纯使用GRU或者LSTM网络模型的分配算法相比，该算法在收敛速度方面加快了5个训练回合，V2V链路间的信道碰撞率和空闲率降低了约27%,平均成功率提升了约10%。

A Channel Allocation Method for Internet of Vehicles Based on GRU and LSTM Hybrid Model

For the problems of channel conflict between Vehicle to Vehicle(V2V) links and low network utility caused by high communication requirements and high mobility in the Internet of Vehicles(IoV),a new channel allocation algorithm for the IoV based on hybrid model of parallel Gated Recurrent Unit(GRU) and Long Short-Term Memory(LSTM) is proposed.This algorithm aims to reduce the V2V links channel collision rate and idle rate,models the channel allocation problem as a distributed deep reinforcement learning problem,makes each V2V link as a single agent,and performs centralized training and distributed execution by maximizing the average reward per episode.In the training process,the hybrid advantages of the short training period of GRU and the high fitting accuracy of LSTM are used to fit the Q function in deep double Q-learning,so that the V2V links can quickly learn and optimize the channel allocation strategy to reuse the Vehicle to Infrastructure(V2I) links channel resources reasonably and maximize network utility.Simulation results show that compared with the allocation algorithm that simply uses the GRU or LSTM network model,the proposed algorithm accelerates the convergence rate by 5 training episodes,reduces the channel collision rate and idle rate between V2V links by about 27%,and increases the average success rate by about 10%.