Synthesis of Analytical Models and Machine Learning Methods for Load Balancing under Strict Response Time Constraints

The subject and purpose of the work. The article is devoted to solving the problem of load balancing in conditions of impatient users and heterogeneous traffic in telecommunication systems. The aim of the work is to develop a conceptual framework that combines analytical methods of queuing theory and modern machine learning approaches to minimize the cumulative penalty, taking into account maintenance delays and the cost of failures. The methods used. The research is based on the M/G/1/K analytical model with impatient users, which makes it possible to evaluate the key indicators of the system. For cases where an analytical solution is impossible or ineffective, the use of (1) time series forecasting to predict load, (2) binary classification to estimate the probability of outflow, and reinforcement learning to optimize the objective function is proposed. The novelty. The difference lies in a systematic approach to combining analytical and ML methods for balancing tasks, as well as taking into account the heterogeneous cost of failures for different traffic classes. A new formalization of the task is proposed through the prism of reinforcement learning. The main results. The concept of an intelligent balancing system has been developed, demonstrating potential advantages over traditional methods. Practical significance. The results can be used in the design of load management systems in autonomous networks.