Multi-Head Attention Reinforcement Learning for Computing Resource Allocation in Flexible Distribution Networks

In a Flexible Distribution Network (FDN), allocating computing resources to minimize latency and energy while guaranteeing task processability is a significant challenge. Heuristic algorithms and single-agent reinforcement learning suffer from the curse of dimensionality, because FDN nodes are highly heterogeneous and the task scale is massive. Multi-agent reinforcement learning (MARL) is a promising alternative, yet existing methods fail to capture the relationship between local observations and the global state and underuse temporal information, which makes them fragile under bursty task arrivals. We propose Attention QMIX (AQMIX), which embeds a state Q-value attention mechanism into QMIX so that multi-head attention explicitly links the global state with local observations, and adds a self-attention state encoder that extracts temporal features to guide the hypernetwork. On the IEEE 33-bus system, a field-derived 13-node scenario and scale-up scenarios with up to 150 nodes, AQMIX completes 99.5–100% of tasks at an on-time rate of at least 99.3%, and lowers energy per completed task by 14–34% relative to QMIX on the 33-bus system and by 47–87% in the scale-up scenarios.