In line with the trend of “native intelligence”, artificial intelligence (AI) will be more deeply integrated into communication networks in the future. Quality of AI service (QoAIS) will become an important factor in measuring the performance of native AI wireless networks. Networks should reasonably allocate multi-dimensional resources to ensure QoAIS for users. Extended Reality (XR) is one of the important application scenarios for future 6G networks. To ensure both the accuracy and latency requirements of users for AI services are met, this paper proposes a resource allocation algorithm called Asynchronous Multi-Agent Deep Deterministic Policy Gradient with Independent State and Action (A-MADDPG-ISA). The proposed algorithm supports agents to use different dimensional state spaces and action spaces; therefore, it enables agents to address different strategy issues separately and makes the algorithm design more flexible. The actions of different agents are executed asynchronously, enabling actions outputted earlier to be transmitted as additional information to other agents. The simulation results show that the proposed algorithm has a 10.41% improvement compared to MADDPG (Multi-Agent Deep Deterministic Policy Gradient). Furthermore, to overcome the limitations of directly applying AI or manual rule-based schemes to real networks, this research establishes a digital twin network (DTN) system and designs pre-validation functionality. The DTN system contributes to better ensuring users’ QoAIS.
Zhou et al. (Tue,) studied this question.