Publications

This paper proposes a reinforcement learning framework that optimizes beamforming codebooks for mmWave and THz MIMO systems using only receive power measurements, eliminating the need for prior channel knowledge and reducing the overhead that hinders highly mobile applications. Through the first comprehensive comparative study of DDPG, TD3, and SAC under realistic conditions including NLoS and hardware impairments, the Soft Actor-Critic algorithm emerges as the top performer, achieving superior beamforming gain and faster convergence across diverse scenarios.

This paper explores using reinforcement learning to train autonomous agents to replicate flocking behaviour in a continuous 2D environment, moving beyond traditional static rule-based approaches. The result is a more adaptable model that deepens our understanding of natural collective motion and opens new possibilities for controlling swarming behaviours across diverse real-world applications.