Online Network Resource Allocation (ONRA) for service provisioning is a fundamental problem in communication networks. As a sequential decision-making under uncertainty problem, it is promising to approach ONRA via Reinforcement Learning (RL). But, RL solutions suffer from the sample complexity issue; i.e., a large number of interactions with the environment are needed to find an efficient policy. This is a barrier to utilize RL for ONRA as on one hand, it is not practical to train the RL agent offline due to lack of information about future requests, and on the other hand, online training in the real network leads to significant performance loss because of the sub-optimal policy during the prolonged learning time. The performance degradation is even higher in non-stationary ONRA where the agent should continually adapt the policy with the changes in service requests. To address this issue, we develop a general resource allocation framework, RADAR, using model-based RL for a class of ONRA problems with known immediate rewards. RADAR improves sample efficiency via exploring the state space in background and exploiting the policy in decisiontime using synthetic samples by the model of the environment, which is trained by real interactions. Applying RADAR on the multi-domain service federation problem to maximize profit via selecting proper domains for service requests deployment, shows its continual learning capability and up to 44% performance improvement w.r.t. the standard model-free RL solution.