Student Pavel Alekseevich Egorov was faced with the task of introducing optimization algorithms into the machine-learning-ui package, designed to solve problems of hydro-aeromechanics and molecular kinetics using machine learning methods. In recent years, the complexity of gas kinetics simulations has increased by several orders of magnitude, but has still not caught up with the complexity of theoretical models. Therefore, in the last five years, optimizations using neural network and regression predictions within numerical methods have been actively studied. machine-learning-ui is one of the tools designed for training such models. An important part of learning most machine learning methods is optimization. Moreover, in many specific cases, basic optimization algorithms are not enough to achieve the required training accuracy.

Pavel Alekseevich's work provides an overview of the AdaMod, MADGRAD, RAdam, Apollo, AdaHessian, LARS and LAMB algorithms. These algorithms have been implemented and implemented in machine-learning-ui. An experimental comparison of the efficiency and accuracy of the algorithms was carried out. A small test was also carried out on the problem of calculating the rate coefficients of energy transitions for a three-component gas mixture.

There are several disadvantages to the work:
The choice of algorithms is quite random. Many algorithms recommended by the supervisor that might be useful in the package, such as SPSA, were not implemented.
No comparison was made with very basic algorithms such as gradient descent.
The work was not carried out particularly evenly and most of the results were obtained in the last two months.
The work lacks a detailed analysis of the results and does not put forward recommendations for the use of algorithms.
I believe that the work can be assessed as “good”, and Pavel Alekseevich Egorov can be awarded a master’s degree.