This paper considers the problem of task scheduling in manufacturing systems with multiple machines operating in parallel. Four approaches to solving this problem are proposed: pure Monte Carlo Tree Search (MCTS), a hybrid MCDDQ agent combining reinforcement learning based on Double Deep Q-Network (DDQN) and Monte Carlo Tree Search (MCTS), an improved MCDDQ-SA agent integrating the Simulated Annealing (SA) algorithm to improve the quality of solutions, and a greedy algorithm (Greedy). A model of the environment is developed that takes into account machine speeds and task durations. A comparative study of the effectiveness of methods based on the makespan (maximum completion time) and idle time metrics is conducted. The results demonstrate that MCDDQ-SA provides the best balance between scheduling quality and computational efficiency due to adaptive exploration of the solution space. Analytical tools for evaluating the dynamics of the algorithms are presented, which emphasizes their applicability to real manufacturing systems. The paper offers new perspectives for the application of hybrid methods in resource management problems.

Keywords: machine learning, Q-learning, deep neural networks, MCTS, DDQN, simulated annealing, scheduling, greedy algorithm

In this experiment, a solver (NEAT) and a simulator (an inverted pendulum cart object) are implemented, where the solver will influence the object in order to keep it in a stable state, i.e. don't let the pendulum fall. The main objective of the experiment is to study the possibility of implementing a simulator of a real physical object and use it to determine the target function of the neuroevolutionary algorithm NEAT. Solving this problem will make it possible to implement controllers based on the NEAT algorithm, capable of controlling real physical objects.

Keywords: machine learning, non-revolutionary algorithms, genetic algorithms, neural networks

The aim of this work is the implementation and comparison of genetic algorithms in the framework of the problem of reinforcement learning for the control of unstable systems. The unstable system will be the CartPole Open AI GYM object, which simulates the balancing of a rod hinged on a cart that moves left and right. The goal is to keep the pole in a vertical position for as long as possible. The control of this object is implemented using two learning methods: the neuroevolutionary algorithm (NEAT) and the multilayer perceptron using genetic algorithms (DEAP).

Keywords: machine learning, non-revolutionary algorithms, genetic algorithms, reinforcement learning, neural networks

The purpose of this work is to implement a system for predicting electricity consumption in food production and to select the most appropriate method for training the forecasting model. In this work, a system was implemented for predicting electricity consumption based on streaming data, which receives them in "real time". The system is implemented on the principle of microservice architecture, where the following were implemented: a service for collecting data from meters, a service for data aggregation and forecasting services. Two forecasting services were implemented: using the classical learning approach based on the ARIMA model and the online learning approach using the HATR online model, the results of which were compared using tests for predicting anomalous values and forecasting under conditions of a change in the data concept, or drift concepts.

Keywords: machine learning, online learning, online model, concept drift, data drift