Dynamic and appropriate resource dimensioning isa crucial issue in cloud computing. As applications go more andmore 24/7, online policies must be sought to balance performancewith the cost of allocated virtual machines. Most industrialapproaches to date use ad hoc manual policies, such as thresholdbasedones. Providing good thresholds proved to be tricky andhard to automatize to fit every application requirement. Researchis being done to apply automatic decision-making approaches,such as reinforcement learning. Yet, they face a lot of problemsto go to the field: having good policies in the early phasesof learning, time for the learning to converge to an optimalpolicy and coping with changes in the application performancebehavior over time. In this paper, we propose to deal with theseproblems using appropriate initialization for the early stages aswell as convergence speedups applied throughout the learningphases and we present our first experimental results for these.We also introduce a performance model change detection onwhich we are currently working to complete the learning processmanagement. Even though some of these proposals were knownin the reinforcement learning field, the key contribution of thispaper is to integrate them in a real cloud controller and toprogram them as an automated workflow.