Abstract : Open RAN (O-RAN) enables multi-vendor interoperability and data-driven RAN control through independently developed xApps, yet concurrent xApps can still issue conflicting actions even when each model is trained and validated offline. This paper proposes a context-aware scheduler in the Near-RT RIC that mitigates xApp conflicts by selecting which pre-trained xApp policies are allowed to update RAN control parameters under the current network context and operator intent. The scheduler is trained offline using Advantage Actor-Critic (A2C) while treating the xApps as immutable components. We study an indirect conflict between a power-allocation xApp and a resource block group (RBG) allocation xApp and evaluate two scheduling policies: (i) dynamic prioritization with action retention and (ii) baseline-augmented coordination that includes deterministic equal-allocation controllers as safe modes. Simulation results in a multi-cell downlink OFDM setting show that uncoordinated concurrent deployment can incur substantial throughput loss relative to single-xApp baselines, whereas the proposed scheduler consistently improves normalized throughput and reduces discarded traffic across diverse traffic and mobility contexts. The baseline-augmented policy yields the best performance by enlarging the scheduler action repertoire and enabling context-dependent composition of learned and baseline decisions. Overall, the results highlight the context-dependent nature of xApp conflicts and demonstrate that coordination can be achieved without joint xApp training, digital-twin action scoring, or post-deployment model modification.
Index terms : O-RAN, Conflict Resolution, A2C, xApp