Learning-based self-adaptation often ignores that switching between configurations is costly. Early in a run the learner also lacks evidence, which leads to unstable behavior. This thesis proposes a controller that optimizes net reward, defined as utility minus reconfiguration cost, and that uses a short, structured warm-up before handing control to a contextual multi-armed bandit. The cost-aware decision model combines fixed and normalized cost terms for booting servers, changing energy use, adjusting a dimmer, and making large configuration jumps. The warm-up selects diverse and nearby candidates, filters unsafe moves with cost guards, and adds a short optimism bonus to untested options. We implement the controller as an external manager in a MAPE-K loop and evaluate it on the SWIM exemplar under trace-driven workload. Across runs the composition improves early and cumulative net reward compared to cost-blind and warm-up-free baselines. It reduces small back-and-forth switches and reaches good behavior sooner. We analyze scalability of candidate generation and show that lazy and approximate variants avoid the quadratic cost of a full distance matrix. The discussion identifies limits, such as sensitivity to cost weights and delayed feedback.