Curiosity-Driven Deep-RL for Adaptive Handover Optimization in O-RAN on the NH-47 Corridor

In recent years, O-RAN has emerged as a major shift in cellular networks, replacing proprietary solutions with open, interoperable, and software-driven designs. It facilitates the integration of AI/ML frameworks to enable adaptive and intelligent network management using Near-RT and Non-RT RIC. One of the critical challenges in such networks is handover management. The traditional methods rely on event-based triggers and limited metrics through hard-coded policies, which often struggle to adapt to dynamic user mobility scenarios. The lack of adaptive policy mechanisms and the dependency on static configurations often hinder seamless mobility management in O-RAN. To overcome these limitations, this paper proposes a Curiosity-Driven Deep Reinforcement Learning framework named C-DDQN for O-RAN-compliant handover control. The agent receives intrinsic rewards based on its exploration by using the Intrinsic Curiosity Module (ICM) integrated with the Double Deep Q-Network (DDQN). The proposed method achieves higher data rates while simultaneously reducing block rate by 10.2%, improving load balancing by 46%, and achieving a 2x improvement in handover success rate on average across both scenarios. These improvements demonstrate the potential of curiosity-driven exploration to enhance the adaptability and efficiency of handover decisions within O-RAN architectures.