1: \begin{abstract}

2: We consider a "fresh" learning-based caching framework for content distribution networks (CDNs) with distributed front-end local caches and a dynamic back-end database. The users prefer the most recent version of the dynamically updated content, while the local caches lack knowledge of item popularity and refresh rates. We first explore scenarios with requests arriving at a local cache following a Poisson process, whereby we prove that the optimal policy features a threshold-based structure. Based on this foundation, we propose a model-based learning algorithm for caching dynamic content. The simulation demonstrates near-optimal cost for Poisson process arrivals and strong performance with limited cache sizes. For more general environments, we present a model-free Reinforcement Learning (RL) based caching policy without prior statistical assumptions. The model-based policy performs well compared to the model-free policy when the variance of interarrival times remains moderate. However, as the variance increases, RL slightly outperforms model-based learning at the cost of longer training times, and higher computational resource consumption. Model-based learning's quick adaptability to environmental changes and its fast convergence rate, makes it a desirable approach for dynamic network environments. This framework offers efficient fresh caching solutions for content distribution networks.

3: \end{abstract}

4: