Learning-based Power Delay Profile Estimation for 5G NR via Advantage Actor-Critic (A2C)

This paper explores to estimate power delay profile (PDP) effectively in order to aid bundle-based channel estimation under 3GPP new radio (NR) specification. The 3GPP radio access network (RAN) working group (WG) 1 adopts that data signals can be bundled every 2 or 4 resource block (RB) with their own precoding. In general, a base station called gNB could decide the precoding with its own algorithm, which is transparent to user equipment (UE). As a result, UE could not estimate PDP in a time domain due to the possibility that precoding is different per bundle in the same symbol. Consequently, this report investigates the PDP estimation via channel correlation across bundles in a frequency domain with help of deep learning. This paper does not pursue the estimated PDP to be close to the ideal one. Instead, we target to minimize the mean square error of channel estimation by using the deep learning technique, called advantage actor-critic (A2C). Both approaches look similar but cannot be guaranteed to be the same since channel estimation in practice is still approximated. The A2C algorithm using the policy gradient keeps improving rewards we defined in this paper iteratively. The proposed scheme can even outperform the performance using ideal PDP by adapting a trained network to channel estimation errors.