Reinforcement Learning for Relay Positioning in Mobile Networks

The relay network performs self-learning through reinforcement learning, where each relay independently learns optimal positioning policies through trial and error without requiring external provision of CSI statistics. The system serves itself by converting the complex task of acquiring CSI statistics into a learning process that automatically adapts to environmental characteristics.

The patent replaces the traditional mechanical approach of explicitly acquiring and processing CSI statistics with a neural network-based reinforcement learning system. The neural network learns positioning policies directly from environmental interactions, substituting the complex information processing mechanism with a learning-based approach that achieves similar or better performance without requiring explicit CSI statistics.

The neural network performs preliminary learning during idle periods or offline phases, accumulating knowledge about the environment and optimal positioning strategies. This preliminary action allows the system to make rapid positioning decisions during actual communication without requiring real-time acquisition of CSI statistics, thus reducing overhead time while maintaining reliability.

The reinforcement learning framework implements continuous feedback loops where the neural network receives feedback from positioning outcomes and channel conditions, continuously refining its policies. This feedback mechanism allows the system to adapt to changing environmental conditions without requiring substantial overhead for explicit statistics collection, as the learning is integrated into the normal operation.