Management and orchestration (MANO) of re-sources by virtual network functions (VNFs) represents one of thekey challenges towards a fully virtualized network architectureas envisaged by 5G standards. Current threshold-based policiesinefficiently over-provision network resources and under-utilizeavailable hardware, incurring high cost for network operators,and consequently, the users. In this work, we present a MANOalgorithm for VNFs allowing a central unit (CU) to learnto autonomously re-configure resources (processing power andstorage), deploy new VNF instances, or offload them to the cloud,depending on the network conditions, available pool of resources,and the VNF requirements, with the goal of minimizing a costfunction that takes into account the economical cost as wellas latency and the quality-of-service (QoS) experienced by theusers. First, we formulate the stochastic resource optimizationproblem as a parameterized action Markov decision process(PAMDP). Then, we propose a solution based on deep reinforce-ment learning (DRL). More precisely, we present a novel RLapproach, called parameterized action twin (PAT) deterministicpolicy gradient, which leverages anactor-critic architecturetolearn to provision resources to the VNFs in an online manner.Finally, we present numerical performance results, and map themto 5G key performance indicators (KPIs). To the best of ourknowledge, this is the first work that considers DRL for MANOof VNFs’ physical resources.