OLRAMT-DEC: Online Learning-Based Resource Allocation for AI Model Training in a Device-Edge-Cloud Continuum

The advancement of applications based on various Deep Neural Network (DNN) architectures, such as Large Language Models (LLMs) has led to an increasing need for the efficient utilization of network resources to meet the strict latency and computational requirements of model retraining. This paper proposes OLRAMT-DEC, an online learning-based device-edge-cloud cooperation framework to address the resource allocation problem for retraining AI models. We leverage the multi-armed bandit (MAB) framework to minimize the retraining latency, while considering time-varying channel conditions. We demonstrate the effectiveness of our approach in adapting to network dynamics through latency and regret minimization under different user loads. Furthermore, we show that our approach outperforms fixed and random allocation baselines in reducing latency toward the optimal benchmark in hindsight.