Towards an Interpretable Fuzzy Approach to Experimental Design

We present in this article an interpretable fuzzy approach to experimental design under constraints that can be used with few data. It is mainly intended (but not limited) to materials science. The goal is to provide experimentalists with an interpretable and explainable algorithm, allowing them to sample optimally the design space. We detail the different steps of our algorithm that consists in recommending the next experiment to perform and building a Sugeno fuzzy rule base. We then present some results on a toy and a real-world datasets. As the method is inspired from Bayesian optimization, we also compare the fuzzy approach to the Bayesian one.