Reinforcement Learning-Based Resource Allocation and Request Routing Optimization Algorithm for Edge AIGC Services

With the explosive growth of artificial intelligence generated content (AIGC) applications, deploying lightweight tiny neural networks (TNNs) at the edge to mitigate the high latency and cost of large cloud models has become an important trend in mobile computing. However, resource heterogeneity in edge environments, the spatiotemporal randomness of user requests, and the complex association problem arising from overlapping base-station coverage challenge traditional static resource allocation methods in meeting multi-dimensional quality of service (QoS) requirements. Existing research often lacks a unified trade-off among model accuracy, service cost, and response latency. To address these challenges, this paper formulates a joint model deployment and request routing multi-objective optimization problem. To handle the high-dimensional, mixed decision space, a dual-timescale decoupled optimization strategy is designed: on the macro timescale, a greedy algorithm based on request statistics periodically performs model deployment to adapt to the long-term traffic distribution and maximize local cache utilization; on the micro timescale, an improved deep deterministic policy gradient (DDPG) algorithm performs real-time dynamic request routing, achieving a multi-dimensional performance trade-off through online environment learning. Simulation experiments based on the real-world EUA trace and calibrated synthetic workloads show that the proposed scheme improves the comprehensive objective and reduces the service cost compared with traditional heuristics and state-of-the-art deep reinforcement learning baselines (TD3 and SAC), while significantly reducing the response latency. Ablation, robustness, and sensitivity studies further confirm the contribution of each component and the stability of the method. The algorithm demonstrates robustness and scalability under high concurrency, indicating its potential for future large-scale edge AIGC services.