针对当前多智能体多用户的MEC环境中的AIGC任务卸载问题，本文以音频语言模型的任务卸载为例，建立了...

In the context of Multi-Agent Multi-User (MEC) environments for AIGC task offloading, this paper takes the task offloading of audio language models as an example and establishes a multi-agent multi-access edge computing system model with the primary goal of maximizing user service quality. Traditional multi-agent reinforcement learning methods often exhibit instability and communication issues when faced with diverse network environments and computational demands. To overcome these challenges, this paper proposes a Multi-Agent Active Inference (MA-AIF) algorithm based on belief-encoded communication. The proposed algorithm extends the TC-AIF algorithm to multi-agent collaborative scenarios, where each agent adopts a communication scheme based on a multi-layer attention mechanism to selectively focus on the belief encoding of other agents, thus obtaining a richer environmental understanding and more accurate state estimation. Furthermore, this paper introduces a value-decomposition-based reward allocation scheme, which leverages a hybrid Q-value network to dynamically adjust the rewards among multiple agents, enabling more coordinated decision-making under different resource conditions and communication environments. Experimental results demonstrate that the MA-AIF algorithm exhibits superior performance and stability across various multi-access edge computing scenarios.

In the single-agent multi-user MEC scenario, this paper addresses the task offloading and resource allocation requirements of video diffusion models by proposing a cloud-edge collaborative offloading architecture that jointly optimizes communication, computation, and caching resources. Building on traditional reinforcement learning research, we introduce the concepts of active inference and temporal consistency, proposing a Temporal Consistency-Based Active Inference (TC-AIF) task offloading strategy optimization algorithm. This algorithm abstracts high-dimensional dynamic scenes and uses temporal-scale reasoning predictions to obtain more meaningful low-dimensional environmental information. The self-supervised temporal consistency method ensures accurate capture of environmental changes, combining active inference with reinforcement learning and designing the optimization objective within the free energy theory framework. To further enhance learning efficiency, we propose a free-energy-based experience replay mechanism, effectively altering the data sampling distribution. Experimental results across multiple scenarios demonstrate that TC-AIF achieves optimal task offloading and computational resource allocation strategies, mitigating the performance impact of exploration-exploitation and data efficiency issues.

In the single-agent multi-user MEC scenario, this paper addresses the task offloading and resource allocation needs of video diffusion models by proposing a cloud-edge collaborative offloading architecture that jointly optimizes communication, computation, and caching resources. Building on traditional reinforcement learning, we introduce active inference and temporal consistency concepts, proposing a Temporal Consistency-Based Active Inference (TC-AIF) task offloading strategy optimization algorithm. By abstracting high-dimensional dynamic scenes and utilizing temporal-scale reasoning predictions, the algorithm captures more meaningful low-dimensional environmental information. A self-supervised temporal consistency method ensures accurate environmental change detection. Combining active inference with reinforcement learning, the optimization target is designed within the free energy theory framework. To enhance learning efficiency, we propose a free-energy-based experience replay mechanism, which alters the data sampling distribution. Experimental results demonstrate that TC-AIF achieves optimal task offloading and resource allocation strategies, mitigating the performance impact of exploration-exploitation and data efficiency issues.

In the multi-agent multi-user MEC environment for AIGC task offloading, this paper uses the task offloading of audio language models as an example to develop a multi-agent multi-access edge computing system model aimed at maximizing user service quality. Traditional multi-agent reinforcement learning methods often exhibit instability and communication issues when facing diverse network environments and computational demands. To address these challenges, we propose a Multi-Agent Active Inference (MA-AIF) algorithm based on belief-encoded communication. The algorithm extends the TC-AIF algorithm to multi-agent collaborative scenarios, where each agent selectively focuses on other agents' belief encodings using a multi-layer attention-based communication scheme to gain richer environmental understanding and more accurate state estimation. Additionally, we introduce a value-decomposition-based reward allocation scheme, using a hybrid Q-value network to dynamically adjust rewards among agents, enabling more coordinated decision-making under varying resource and communication conditions. Experimental results show that MA-AIF exhibits excellent performance and stability across multiple multi-access edge computing scenarios.