Résumé
Fuzzy arithmetic is useful to handle uncertainty in data. They usually rely on costly interval arithmetic based on lower-upper representation. In some cases, fuzzy numbers are symmetrical. We will show how we can exploit this property to halve the amount of operations and memory needed and increase accuracy as well, thanks to the midpoint-increment representation format. We will describe how we implemented this format into a fuzzy arithmetic library, specifically tuned to run on Graphic Processing Units (GPU). The results of a series of tests using compute-bound and memory-bound benchmarks show that the proposed format provides a performance gain of two to twenty over the lower-upper format.