1: \begin{abstract}

2: To ensure safety in teleoperated driving scenarios, 

3: communication between vehicles and remote drivers must satisfy strict latency and reliability  requirements.

4: In this context, \gls{pqos} was investigated as a tool to predict unanticipated degradation of the \gls{qos}, and allow the network to react accordingly. %in order to guarantee a target performance. 

5: %In previous studies, \gls{rl} has been identified as a possible means to perform accurate predictions.

6: In this work, we design a \mbox{\gls{rl}} agent to implement \gls{pqos} in vehicular networks. To do so, based on data gathered at the \gls{ran} and/or the end vehicles, as well as \gls{qos} predictions, our framework is able to identify the optimal level of compression to send automotive data under low latency and reliability constraints.

7: %To do so, based on data gathered from the \gls{ran} and/or end vehicles, and identifies the optimal level of compression to send automotive data 

8: We consider different learning schemes, including centralized, fully-distributed, and federated learning. We demonstrate via ns-3 simulations that, while centralized learning generally outperforms any other solution, decentralized learning, and especially federated learning, offers a good trade-off between convergence time and reliability, with positive implications in terms of privacy and complexity. %needed for the \gls{rl} agent to converge.

9: \end{abstract}

10: