A Multi-Agent Framework for AI-Generated Content Caching in Internet of Vehicles Networks

As Internet of Vehicles (IoV) networks in urban autonomous driving environments increasingly incorporate generative artificial intelligence (GAI) for real-time monitoring and management of autonomous vehicle fleets, ensuring timely and efficient delivery of artificial intelligence-generated content (AIGC) is essential for maintaining black seamless traffic operations and mobility services. This paper tackles the multi-objective challenge of minimizing age of information (AoI), ensuring high inference accuracy, and optimizing energy use under constrained edge computing resources by introducing a novel multi-agent reinforcement learning framework for multi-objective tasks (MARMOT). Leveraging the multi-agent twin delayed deep deterministic policy gradient (MA-TD3) algorithm, MARMOT jointly optimizes content caching, uplink power allocation, and GAI inference placement in multi-access edge computing enabled IoV networks. The proposed framework integrates autoencoder-based state compression with a multi-head critic architecture, enabling scalable and efficient learning among edge nodes. Simulation results confirm that MARMOT significantly outperforms state-of-the-art baselines such as multi-agent deep deterministic policy gradient and standard MA-TD3, achieving reduced AoI and AoI violation probability, superior inference accuracy, and enhanced energy efficiency across black diverse autonomous vehicle fleet deployment scenarios.