Introduction to deep machine learning based on physically defined data - Hamiltonian / Lagrangian (physically informed) neural networks

Deep machine learning (GMO) has achieved significant results in many tasks with large amounts of data and generalization in close proximity to the training data. For many important real-world applications, these requirements are not feasible, and additional prior knowledge in the subject area is required to overcome emerging problems and "pathological" logical paradoxes. In particular, studying physical models for model-based control requires reliable extrapolation from fewer samples, which are often collected online, and model errors can lead to serious damage to the system. We consider Benchmarks as examples of the machine learning (ML) effectiveness method taking into account physics (Lagrangian and Hamiltonian neural networks) when studying the dynamics model in the state space of an autonomous control object.