Model predictive complex system control from observational and interventional data.

Area of Science:

• Complex Systems Science
• Control Theory
• Machine Learning

Background:

• Complex systems exhibit emergent behaviors, making data-driven modeling and control crucial.
• Traditional control methods struggle with high intervention costs and limited intervention data.
• Abundant observational data is often available, but direct intervention is expensive.

Purpose of the Study:

• To develop a novel framework for controlling complex systems with minimal online interventions.
• To address challenges in high-dimensional state-action spaces within complex systems.
• To leverage abundant observational data for effective system control.

Main Methods:

• Introduced a two-stage model predictive complex system control framework.
• Employed offline pre-training using observational data to model system dynamics.
• Utilized online fine-tuning with a variant of model predictive control for interventions.
• Developed action-extended graph neural networks to model the Markov decision process.
• Designed a hierarchical action space for efficient learning of intervention strategies.

Main Results:

• The proposed framework demonstrated strong performance in Boids, Kuramoto, and SIS metapopulation environments.
• Achieved accelerated convergence and robust generalization capabilities.
• Significantly reduced intervention costs compared to baseline algorithms.
• Effectively handled high-dimensional state-action spaces in complex systems.

Conclusions:

• The two-stage framework offers an effective solution for controlling complex systems with limited intervention data.
• Action-extended graph neural networks and hierarchical action spaces are key innovations for this problem.
• This approach holds promise for real-world applications requiring efficient complex system control.