Deep Reinforcement Learning Based Fast Initial Access For mmWave Based User-centric Systems.

User-centric (UC) systems face a complex problem of initial access (IA) beam discovery due to a large number of distributed access points (APs). The use of millimeter-wave (mmWave) frequencies, further exacerbates the IA problem due to its shorter channel coherence time and higher blockage susceptibility compared to the lower frequency bands. In this paper, we provide a novel deep reinforcement learning-based algorithm called RapidIA to efficiently perform IA in mmWave based UC systems. The key advantage of RapidIA is that it can predict the IA beam steering vector using just one reference signal from the user. The policy of the deep reinforcement learning model in RapidIA learns the mapping between received reference signals and the beam codebook. The mmWave channel and the simulation environment are modeled using a ray-tracing approach. Ray-tracing based simulations show that the RapidIA algorithm consistently achieves a beam prediction accuracy level of 100%. With respect to beam discovery delay, RapidIA shows a considerable improvement compared to other beam sweeping based IA systems.