Boosting in-transit entertainment: deep reinforcement learning for intelligent multimedia caching in bus networks

Multimedia content delivery in advanced networks faces exponential growth in data volumes, rendering existing solutions obsolete. This research investigates deep reinforcement learning (DRL) for autonomous optimization without extensive datasets. The work analyzes two prominent DRL algorithms, i.e., Dueling Deep Q-Network (DDQN) and Deep Q-Network (DQN) for multimedia delivery in simulated bus networks. DDQN utilizes a novel “dueling” architecture to estimate state value and action advantages, accelerating learning separately. DQN employs deep neural networks to approximate optimal policies. The environment simulates urban buses with passenger file requests and cache sizes modeled on actual data. Comparative analysis evaluates cumulative rewards and losses over 1500 training episodes to analyze learning efficiency, stability, and performance. Results demonstrate DDQN’s superior convergence and 32% higher cumulative rewards than DQN. However, DQN showed potential for gains over successive runs despite inconsistencies. It establishes DRL’s promise for automated decision-making while revealing enhancements to improve DQN. Further research should evaluate generalizability across problem domains, investigate hybrid models, and test physical systems. DDQN emerged as the most efficient algorithm, highlighting DRL’s potential to enable intelligent agents that optimize multimedia delivery.