Managing Conflicts Among Black-Box RAN Apps via Multi-Fidelity Game-Theoretic Optimization

Modern open and softwarized networks---such as O-RAN telecom systems---host independently developed applications, known as xApps, with distinct and potentially conflicting objectives. Coordinating their behavior to ensure stable operation is challenging, especially when each application's utility is only accessible via costly black-box evaluations. In this work, we consider a centralized controller that suggests joint configurations to multiple apps, modeled as strategic players, with the goal of aligning their incentives toward a stable outcome. This setting is modeled as a Stackelberg game in which the central controller (leader) lacks analytical forms of the players' utility functions, and must learn them through sequential, multi-fidelity queries. We propose {MF-UCB-PNE}, a novel multi-fidelity Bayesian optimization strategy that efficiently approximates a pure Nash equilibrium (PNE) under a limited query budget. MF-UCB-PNE balances exploration of cheap, lower-fidelity utility observations with exploitation of high-fidelity evaluations, enabling convergence to incentive-compatible configurations. We provide theoretical guarantees in terms of no-regret learning of equilibrium, as well as empirical results on wireless networking problems, demonstrating that MF-UCB-PNE identifies high-quality equilibrium solutions under strict cost budgets.