Reinforcement Learning Based Power-Optimal Usage of Beamforming Antenna Array for Multi-Way Wireless Communication in Vehicular Traffic Environments

There has been recent work on the design of antenna arrays for beamforming in dynamic evolving environments such as in vehicle-to-vehicle communication systems. A key problem is that of determining how to optimally use a large antenna array to communicate with multiple spatially located vehicles in dynamically changing channel conditions with minimal co-channel interference while minimizing overall power consumption of the wireless system. We envision disjoint subsets of antennas in the array being used to direct beams concurrently to different vehicles. The number of antennas, gain and phase of each RF-chain driving an antenna are optimized dynamically using a constrained quadratic cost formulation encompassing channel quality, interference and power consumption. This quadratic optimization problem is solved using behavior constrained bandit algorithm, a reinforcement learning based technique. A gaussian kernel is used to perform data clustering of vehicle environment and resulting solutions, allowing quick bootstrapping of the bandit solver to find optimal array configurations in real-time vehicle environments. Simulation studies prove the viability of the proposed scheme.