Deep Q-Networks Allocating Massive MIMO Phased Array Beams Between Mobile Users and Radar Targets

Neural network policy agents scheduling phased array beam directions at sub-millisecond intervals to maximize joint communication and radar performance.

Deep Q-Network State and Reward Formulations

ISACOrchestrator models beam scheduling as a Markov Decision Process (MDP). Neural policy networks observe user channel states, vehicle kinematics, and queue lengths, selecting beam steering vectors that maximize joint reward functions.

Hardware-Accelerated Neural Inferencing

Trained reinforcement learning policies execute on embedded TensorRT and FPGA accelerators in under 250 microseconds, keeping pace with rapid 1 kHz 6G radio frame slots.

Deep Q-Network beam scheduling executing in under 250 microseconds
Simultaneous tracking of up to 64 independent communication and radar targets
Dynamic beamwidth modulation matching target velocity and distance
Zero training stability collapse through double dueling DQN architectures

Voir les domaines