With increasingly complex urban dynamics, as well as increasing demands for the sustainability of urban mobility and introduction of cognitive technologies into transport infrastructure, the paper proposes a dual-loop method for intelligent traffic light control based on reinforcement learning and phase strategy distillation procedures. The first level implements real-time control through an RL-agent, while the second one generates backup hourly plans based on statistics of its behavior. The method is based on a system-discrete model taking into account stochastic traffic parameters and permissible control constraints. The simulation conducted in SUMO for a real intersection demonstrates a significant reduction in average transport delay compared to classical control, confirming the efficiency, sustainability and scalability of the approach. The obtained results substantiate the possibility of practical implementation of the model within the framework of intelligent transport systems of large cities and for laying the engineering foundation for hybrid urban mobility management architectures.