Exploring KPI Trade-offs in O-RAN Conflict Resolution with Deep Reinforcement Learning

Abstract

Open-Radio Access Network (ORAN) enables independently developed xApps, software plug-ins, to optimize different radio objectives through the Near-Real-Time RAN Intelligent Controller. Their simultaneous operation can, however, generate conflicting actions and trade-offs among network Key Performance Indicators (KPIs), operational cost, and user Quality of Experience (QoE). We investigate Deep Reinforcement Learning (DRL) as an adaptive conflict mediator that observes xApp proposals and learns continuous RAN-parameter adaptations. Its objective jointly accounts for QoE, network cost, and configuration safety. In an integrated multi-xApp scenario, the learned mediator reduces normalized total cost by approximately 40% compared with uncoordinated xApps, mainly by reducing QoE degradation while preserving admission success and fairness.