Interference Cancellation And Resource Allocation For Full-duplex Systems Based On Machine Learning

The increasing demand on high communication transmission rates and system capacity puts great pressure on limited spectrum resource.There exist two solutions.One is exploring more frequency resource and the other is improving spectrum efficiency.The full-duplex(FD)technique enables devices to communicate in the same frequency band at the same time and thus theoretically doubles the frequency efficiency,which has a broad prospection in 5G and future communication systems.At present,the application of FD technique confronts two major challenges.First,simultaneous communication in the same frequency band brings severe selfinterference(SI)and the cancellation of SI is the precondition of FD's application.Second,the solve of system resource optimization problem guarantees the performance of FD in practical systems.For the cancellation of self-interference problem,the traditional least-squares(LS)algorithm can estimate the interference signal accurately but with high computational complexity.Based on this algorithm,we apply the deep unfolding method in machine learning and propose a SI cancellation algorithm by unfolding the iterations of a projected gradient descent algorithm.We compare the performance and complexity of LS algorithm and the proposed algorithm.The simulation results show that the proposed algorithm can achieve good performance while maintaining a low computational complexity.For the resource optimization problem in FD communication system,we conduct our research in two aspects.One is the user pairing problem in FD communication systems.Generally,the macro base station(BS)can work in FD mode while the users only work in half-duplex(HD)mode.To maximize system throughput,it is important to pair users properly.This problem is usually formulated as a non-deterministic polynomial(NP)mixed integer programming problem.In this paper,we propose a user pairing algorithm based on reinforcement learning.The BS can learn optimal paring policy through interacting with the environment.Compared to the traditional user pairing algorithm,the proposed algorithm can achieve nearly the same performance with a much lower computational complexity.Furthermore,when the number of users in the environment changes,the trained model can maintain performance without retraining.The other is the power allocation problem in FD communication system.Due to the simultaneous transmission in the same frequency band,not only the fairness of downlink power allocation should be considered but also the co-channel interference.Besides,the restrictions include the total power of BS and the requirements of users.This problem is usually formulated as a non-convex optimization problem.In this paper,we propose a power allocation algorithm based on reinforcement learning.During the interaction with environment,the BS can learn a power allocation policy.The simulation results show that the proposed algorithm can achieve nearly the same performance as the traditional algorithm but with a lower computational complexity.