HyMAS (Hybrid Multi-Agent System) is an innovative production-planning framework developed to support autonomous and dynamic decision-making in complex steel manufacturing environments. Implemented within an academic agent framework and integrated with the industrial PSImetals product, HyMAS combines intelligent agents, digital twins, and machine learning to optimize production workflows in real time, reacting to unforeseen events and disruptions.
Instead of deploying the system in a real steel facility, we created a virtual steel factory based on the process structure of a thin steel producer. This simulated environment replicated realistic production flows and included typical operational events and disturbances to validate HyMAS under near-real-world conditions.
Challenges
Steel production involves high energy consumption, strict delivery timelines, and vulnerability to disruptions such as equipment failures or material quality issues. Traditional centralized planning systems are often rigid, making them slow and ineffective in responding to unforeseen events. This lack of adaptability leads to efficiency losses and missed delivery targets.
Objectives
Solution Approach
At the core of HyMAS is a hierarchical, modular architecture composed of autonomous agents, each responsible for a distinct level of production planning. These include:
-
Resource Agents (RA) for individual machines or lines
-
Process Agents (PA) for entire processing stages
-
Process Chain Agents (PCA) for coordinating sequential operations across stages
Each agent is equipped with embedded optimization models based on mathematical algorithms, including heuristics and exact methods. These models vary in complexity—some deliver rapid, approximate solutions using lightweight heuristics, while others require more time to compute high-quality results using advanced optimization techniques.
To address the variable response time required in different operational contexts, the HyMAS architecture is designed to produce preliminary results quickly, especially when fast reactions are critical. Initial decisions are based on fast heuristics to maintain responsiveness. Subsequently, more sophisticated models can refine the solution as additional time becomes available, allowing continuous improvement of the plan.
Agents operate in tight cooperation, sharing intermediate results and depending on each other’s inputs to maintain overall system coherence. This cooperative behavior ensures that local decisions align with global objectives and that the system can respond effectively under real-time constraints.
A key feature is the planner’s ability to monitor evolving solutions, as the system updates in real time. Depending on the disruption or event, available reaction time may range from seconds to minutes. By providing early feasible solutions and refining them over time, HyMAS supports agile and informed human oversight without sacrificing speed or quality.
Result
The HyMAS system was validated in a simulated production environment configured to model typical material flows along with a wide range of operational events and disturbances—such as machine breakdowns, quality deviations, inventory bottlenecks, and urgent order changes.
Through this simulation, HyMAS demonstrated a significant improvement in planning responsiveness and adaptability. The system's layered optimization approach together with agent architecture enabled it to quickly deliver initial, feasible plans during disruptions and progressively refine them over time. This staged response allowed planners to visualize the current best-known schedule at any moment and supported decision-making even under tight time constraints.
Achievements
Quantitative evaluations of the simulation runs showed that, compared to a conventional static planning approach, HyMAS achieved:
- Up to 18% increase in operational efficiency,
- 8% improvement in on-time delivery performance,
- Substantially faster rescheduling, with meaningful solutions generated in seconds.
Conclusion
HyMAS demonstrates how a hybrid multi-agent architecture, combining real-time heuristics and advanced optimization, can bring agility and resilience to steel production planning. By simulating realistic scenarios, the system proved its ability to react quickly, refine plans over time, and support human planners with transparent, adaptive decision-making—marking a strong step toward autonomous production control in complex industrial environments.
