A Multi-Agent Graph-Attentive Framework for Predictive Expert Task Scheduling

Efficient scheduling of complex expert tasks is vital for optimizing service delivery and maintaining balanced workloads. Traditional methods often struggle to adapt to dynamic task demands and expert availability. To address this, we propose the Multi-Agent Scheduling and Prediction Network (MASP-Net), which integrates multi-agent reinforcement learning (MARL), graph attention mechanisms, and temporal prediction modules to improve expert task allocation. MASP-Net adopts a shared global critic and local actor structure, enabling agents to collaboratively learn optimal assignments. The graph attention module models task-expert dependencies, while a GRU-based predictor estimates task durations to guide decisions. Data preprocessing constructs skill-task compatibility matrices and bipartite graphs to support graph learning. The loss function combines actor-critic targets with time prediction errors, and uses gradient clipping and experience replay to tackle non-stationarity and sparse rewards. MASP-Net consistently outperforms DQN, DDPG, and rule-based baselines across key metrics, confirming the synergy of its components.