Real world problems often require complex modeling and computation efforts to be effectively addressed. Relying solely on data-driven approaches without integrating physics-based models can result in limited predictive capabilities. Even advanced techniques such as deep learning may be impractical for decision-makers due to the lack of data and challenges in justifying and explaining results. In this paper, we propose INFERNO (INference Framework for Efficient Rule-based kNOwledge generation), a framework designed to address these challenges. Our methodology integrates derivative-free optimization, bipartite network clustering, and a novel procedure to derive explainable inference rules. These rules enable decision-makers to easily identify high-quality solutions. We introduce a new metric, called Price of Explainability (PoX), to quantify the trade-off between quality and explainability.
The framework was validated on two building design problems. Results show that INFERNO achieves PoX values that are, overall, 4.5 to 10 times lower than those of a classification tree.