To support ultra-reliable low latency communication (URLLC) services in fifth-generation mobile networks, short-packet transmission is essential. However, due to the limited packet size, errors cannot be reduced to arbitrarily low levels for a given coding rate as for conventional communication systems covered by the Shannon theory. In this paper, we consider URLLC in dual-hop decode-and-forward relaying networks where the channel in each hop varies fast. A simple but efficient optimization of the block lengths is performed to minimize the block error rate (BLER) of the proposed system. In particular, we deploy machine learning models using the linear regression and normalized method to determine the optimal fraction of channel uses for the transmission over each hop. Numerical results show that the BLER of the consider relaying system with optimal block lengths for each hop based on the machine learning model outperforms conventional relaying systems with equal block lengths. © 2022 IEEE.