Abstract: In future mobile networks, such as 5G networks, network slicing will be a promising technology to provide customizing services for different users with different transmission requirements. According to the dynamic network state in slice based mobile networks, users need to make accessing slice handoff periodically for improving the transmission performance. However, in a multi-user networks, the accessing choice of a user changes the amount of available transmission resources in the system, which impacts the accessing choices of other users. Thus, in this paper, we model the multi-user handoff problem in slice based mobile networks as a multi-agent random game. Then, we use multi-agent reinforcement learning (MARL) to solve this game, and propose a multi-user accessing handoff algorithm based on distributed MARL method. The numerical results validate the performance of our proposed multi-user accessing handoff algorithm in slice based mobile networks.