Abstrakt
Machine learning has already proven itself in many scientific domains, most notably by improving classifications or value-predictions (regression). However, its wide deployment in the field of adaptive systems has yet to come.
There are several reasons for this cautious adoption, such as the lack of sufficient data for training the models or reluctance to incorporate black-box models into the adaptation processes of existing systems. To gap these limitations, we are often required to perform detailed simulations to generate training data or verify the feasibility of adopting ML models in our systems.
We present the ML-DEECo framework that should simplify the design of ML-enabled simulations, which removes repetitive code such as gathering specific data from the simulation, using these data to train models, and applying these models in the simulations as predictors. We also provide two case studies as an example and an evaluation of the usability of the proposed framework.