Federated Multimodal Bi-LSTM for Agentic Handover in 6G Heterogeneous Networks

In 6G Heterogeneous Networks (HetNets), conventional handover mechanisms such as 3GPP Event A3 rely on single-metric thresholds and static parameters. These reactive schemes are often inadequate in dense and highly dynamic deployments, leading to handover failures, pingpong effects, and degraded user experience. This article presents a tutorial overview of an agentic handover framework in which the user equipment (UE) acts as an intelligent agent that predicts its future motion and radio conditions. The framework combines multimodal sequence prediction based on a federated multi-task bidirectional long short-term memory (Bi-LSTM) model with a prediction-driven handover controller. The model jointly forecasts UE trajectory and multi-cell signal indicators while preserving data privacy through federated learning. On top of these predictions, a cosine-similarity-based base station selection strategy and a dynamic Time-to-Trigger (TTT) adaptation scheme are used to reduce unnecessary handovers and ping-pong events. Simulation results in representative 6G HetNet scenarios show that the proposed approach suppresses ping-pong probability to approximately 12–21% (representing a relative reduction of up to 63% compared with classical Event A3/A5 handover strategies), while remaining lightweight enough for edge deployment.