Machine learning based on a generative adversarial tri-model.

Area of Science:

• Engineering
• Machine Learning
• Physics-informed AI

Background:

• Pure machine learning models often violate fundamental physics laws in engineering applications.
• Integrating analytical knowledge into neural networks is crucial for reliable predictions.

Purpose of the Study:

• To propose a novel machine learning paradigm, the generative adversarial tri-model (GAT), for incorporating analytical knowledge into neural networks.
• To address the limitations of pure machine learning models in adhering to physical governing laws.
• To demonstrate the reliability and applicability of the GAT method in solving complex problems.

Main Methods:

• Developed a generative adversarial tri-model (GAT) utilizing a positive-sum game strategy.
• Implemented the GAT method to solve ordinary differential equation (ODE) problems with diverse constraints.
• Proved a strict error bound for initial-constraint problems to ensure reliability.

Main Results:

• Successfully applied the GAT method to solve ODE problems with various constraints.
• Achieved a proven strict error bound for initial-constraint problems, confirming reliability.
• Demonstrated the GAT method's effectiveness in a real-world human body oscillation recovery problem using balance sensor data.
• Validated through human experiments, showing significant improvements in human balancing evaluation.

Conclusions:

• The generative adversarial tri-model (GAT) is a useful and reliable method for integrating analytical knowledge into neural networks.
• The GAT method shows great potential for wider applications and adaptations in various scientific and engineering fields.
• The approach successfully addresses the challenge of ensuring machine learning models obey fundamental physical laws.