Abstract: Disclosed is a multi-base station queued preambles allocation method based on collaboration between multiple agent, which provides a non-competitive preambles allocation method in a scenario of multiple base stations and multiple cells, so as to solve the congestion problem caused during random access of massive MTCDs. The devices are queued to select the preambles based on deep reinforcement learning, and a training method based on federal learning is applied, thus effectively solving the congestion problem caused by competitive access. First, the newly accessed devices are grouped and the level of priority is set according to a delay tolerance time. Then, the device are reasonably allocated to idle queues by means of a multi-agent reinforcement learning algorithm. Finally, by means of a federal training method, the neural network of each agent is synchronously optimized by means of average optimization of neural network gradients, thus completing preambles allocation to each agent.