Computer Science · Computational Theory and Mathematics
Error Detection for Floating-Point Program via Branch and Bound Method
See on the mapAbstract
It is well-known that writing an error-free floating-point program is very difficult. Thus, detecting unacceptable errors of a floating-point program is important. In this paper, we develop a system named SpaceAED. The main function of this system is to automatically detect unacceptable errors of a floating-point program written in C programming language. The key insight of this work is to use interval arithmetic in conjunction with branch and bound technique. The implementation of SpaceAED is to rewrite a floating-point program to one that can run on interval arithmetic, and then use branch and bound technique to find all inputs that can trigger unacceptable errors. We choose a great many of functions in GNU Scientific Library (GSL) to test SpaceAED, including matrix computations and evaluation of special functions etc. Numerical results show that SpaceAED is available for accurately detecting unacceptable error-triggering inputs of numerical functions.