V2N Service Scaling with Deep Reinforcement Learning

In this paper we propose to vertically scale V2N services using DDPG-DOD, a DDPG agent equipped with a nonparametric
discretization method that is designed to capture the structure of the scaling decisions and learn discrete actions in a
continuous fashion – thus avoiding the action-space explosion. Employing a real-world vehicular trace dataset, we show that
DDPG-DOD outperforms state of the art solutions in terms of (i) operational cost as it minimizes the number of active
CPUs, (ii) performance; increasing the long-term reward is an indicator of reduced backlog and thus processing delay, and
(iii) flexibility in scaling resources as DDPG-DOD performs robustly independently of the size of the action space.