Multi-objective Resource Allocation Based on Deep Reinforcement Learning in HetNets

Considering a two-tier HetNet adopting orthogonal frequency division multiple access (OFDMA) in downlink, we propose a multi-agent deep reinforcement learning (MA-DRL) scheme to perform channel assignment and power allocation. The goal is to jointly optimize the system capacity and power consumption. In the proposed scheme, the macro base station (MBS) and pico base stations (PBSs) play the role of agents. Each agent consists of a deep
$Q$
network (DQN) and a deep deterministic policy gradient (DDPG) network performing channel assignment and power allocation in serial. Simulation results show that the proposed algorithm converges steadily and can be used to improve the system capacity and reduce the power consumption effectively.