The invention discloses a heterogeneous network resource allocation method based on reinforcement learning. Firstly, a DNN framework is deployed in each base station. The framework is based on the ADMM algorithm, and the channel information is regarded as the weight of the network; according to the data obtained by the base station, that is The current user association information and average interference power give the best resource allocation strategy in the current state; each base station is regarded as an independent subject, and the state of the base station is used as a modeling environment; several agents observe the same heterogeneous network environment and Take actions, and at the same time, the agents communicate with each other through the rewards of the environment; the agents adjust the policy according to the rewards; the resource allocation method provided by the present invention is based on the deep learning network, and the resource allocation plan can be given without all CSI information, while considering the spectrum efficiency , setting the spectral efficiency function as the reward of the agent can guarantee the spectral efficiency while ensuring the system throughput.