Abstract
This paper proposes a novel machine learning paradigm called the generative adversarial tri-model (GAT) to incorporate analytical knowledge into neural networks through a unique positive-sum game strategy. The motivation is to solve the problem that pure machine learning models fail to obey the fundamental governing laws of physics in engineering fields. The GAT method is successfully implemented to solve ODE (ordinary differential equation) problems with various constraints. A strict error bound is proven for initial-constraint problems, which certifies its reliability. The real-world significance of the GAT method is reflected by its application to a human body oscillation recovery problem, based on balance sensor measurements, which is critical for human balancing evaluation, yet unresolved after massive precedent research work. Further human experiment results prove the effectiveness of the GAT method. Both theoretical and experimental studies demonstrate that the GAT method is useful and reliable. It envisions great scalability for wider applications and adaptions prospect.