Distributed Multi-Expert Forecasting System for Non-Stationary Processes

This paper explores the challenge of predicting complex dynamic systems that exhibit oscillatory non-periodic processes and non-stationary random components. Such systems are prevalent in gas-dynamic and thermodynamic media, as well as in economics and social psychology. The paper proposes a distributed two-level computational scheme with independent intellectual agents based on diverse predictive algorithms and a Bayesian supervisor to create a final forecast. The implementation of a distributed system with a Bayesian supervisor can improve the forecast's overall quality by approximately 8.5% compared to a linear extrapolator and 15.7% compared to a quadratic extrapolator. Results show the potential of this system, but the proposed model has limitations and requires the utilization of more advanced predictor agents and interagent communication technologies.