Automated Online Experiment-Driven Adaptation–Mechanics and Cost Aspects

As modern software-intensive systems become larger, more complex, and more customizable, it is desirable to optimize their functionality by runtime adaptations. However, in most cases it is infeasible to fully model and predict their behavior in advance, which is a classical requirement of runtime self-adaptation. To address this problem, we propose their self-adaptation based on a sequence of online experiments carried out in a production environment. The key idea is to evaluate each experiment by data analysis and determine the next potential experiment via an optimization strategy. The feasibility of the approach is illustrated on a use case devoted to online self-adaptation of traffic navigation where Bayesian optimization, grid search, and local search are employed as the optimization strategies. Furthermore, the cost of the experiments is discussed and three key cost components are examined-time cost, adaptation cost, and endurability cost.