Validation of social agent training: synthesis of reinforcement learning and evolutionary optimization methods

The article proposes a time series forecasting model designed for unstable and partially observable environments. Unlike traditional approaches, the developed FELAR architecture combines local agent learning with reward adjustment based on collective characteristics (trust, reputation, influence), alongside global evolutionary adaptation of strategies. The proposed model operates in a distributed multi-agent environment, enabling both local adaptive behavior and global strategy evolution. A set of experiments on publicly available time series datasets (urban traffic, transformer temperatures, electricity consumption) confirms the model’s high forecasting accuracy and robustness to concept drift. The article details the agent architecture, algorithmic loop, experimental setup, and computational efficiency of the approach. The paper highlights key advantages of the approach, including robustness to concept drift, real-time adaptability, and low computational overhead.