With the advent of modern multi-core architectures, the problem of scheduling complex real-time applications has become one of the major challenges in developing safety-critical systems. In this context, existing solutions typically rely on greedy, hand-crafted heuristics and/or pessimistic tests that fail to meet the expected high performance and timing guarantees. This paper advocates an efficient learning-based scheduling framework that automatically generates an adaptive policy for a given workload and platform, thereby mitigating these limitations.