Reinforcement learning based mobility load balancing in cellular networks: a two-layered approach

Load balancing in cellular networks has been gaining great importance with the increasing number of users. When there are too many users in a small area, some of the cells may become overloaded and the full capacity of the network performance cannot be used. This can be prevented by adjusting Cell Individual Offset (CIO) values which controls the thresholds between cells to start a handover procedure of the users. However, adjusting the CIO values to the current state of the network does not work well when too many users are in motion. In this study, we propose a two-layered load balancing solution. At the first layer, a centralized agent collects the latest handover numbers from the network and predicts the number of users that will be connected to each cell in the next time frame. The second layer uses the predicted numbers and adjusts CIO values by using reinforcement learning techniques. We compared our solution with a decentralized system without the load prediction layer and a system with constant CIO values. The simulation results show that our two-layered approach achieves a more balanced distribution for the users.