The increasing complexity of modern applications has motivated the need to automate their management functions. The applications are then able to manage themselves and meet their SLA requirements by means of autonomic MAPE-K loops based on predefined policies. However, the common use of fixed and hand-coded policies, known for being knowledge-intensive, is inadequate to dynamically changing contexts. Autonomic management should be dynamically adaptive to learn appropriate policies at runtime. Towards this direction, we propose to provide autonomic systems with learning abilities to render the decision making self-adaptive. In this paper, we propose to enrich the decision making process of an autonomic MAPE-K loop with a learning-based approach. We demonstrate the usage of learning techniques as building blocks of sophisticated and better performing autonomic systems. We have illustrated our approach with a real-world application example. The experimental results have shown a dynamic adjustment to a changing context in a shorter time as compared to existing approaches. They have also shown less frequent time spent in SLA violations during the learning phase. The approach converges faster and demonstrates higher efficiency and better learning performance