An AI-Driven Adaptive Framework for Real-Time Performance Optimization in Unity-Based Mobile Games

Abstract

Sustained mobile gameplay is a dynamic performance-management problem because scene complexity, memory pressure, and thermal state change during play. This study proposes a safety-constrained AI framework for Unity-based mobile games that combines runtime telemetry, deterministic thermal safeguards, and a lightweight contextual multi-armed bandit for reversible fidelity decisions. The framework is grounded in authentic public data rather than simulated controller gains. A verified 30-run Unity Energy Benchmark file for Fantasy Kingdom was re-analysed. Mean frame rate was 116.674 FPS (SD = 0.062; 95% CI [116.651, 116.697]), whereas combined CPU and GPU power averaged 11.125 W (SD = 1.427; 95% CI [10.592, 11.657]). Relative variability was 0.053% for FPS and 12.83% for combined power, showing that a stable FPS average can conceal substantial resource variation. A separate University of Southampton mobile thermal dataset provides supporting evidence on throttling, jank, power, and skin temperature. Based on these observations, the proposed controller uses frame timing, memory, scene and quality context, and Android thermal headroom to select safe device-specific actions. The paper contributes a reproducible architecture, reward and action formulation, data-provenance procedure, instrumentation design, and Android validation protocol. Received: 11 September 2026 / Accepted: 28 October 2026 / Published: November 2026