New AI Framework Boosts Multi-Agent Coordination in Dynamic Manufacturing.

Dynamic manufacturing environments frequently face disturbances like machine failures or urgent job arrivals, requiring multi-agent systems to adapt quickly. Traditional AI approaches often treat each disturbance as a new problem, discarding valuable past coordination experience. This new research introduces the Graph-Structured Experiential Memory (GSEM) framework to address this limitation. GSEM works by encoding past coordination episodes as heterogeneous relational graphs, capturing critical information such as task dependencies, machine states, and how agents collaborated. When a new disturbance occurs, a graph neural network efficiently retrieves structurally similar past episodes. This enables the system to adapt its policies using learned experience rather than starting from scratch. Experiments on flexible job-shop scheduling benchmarks demonstrated that GSEM significantly reduces makespan and adaptation time compared to existing memory-augmented baselines. Its advantages become more pronounced with higher disturbance frequencies, highlighting its potential for robust, adaptive manufacturing operations.