Generating diverse software versions with genetic programming: an experimental study

Software fault-tolerance schemes often employ multiple software versions developed to meet the same specification. If the versions fail independently of each other, they can be combined to give high levels of reliability. Although design diversity is a means to develop these versions, it has been questioned because it increases development costs and because reliability gains are limited by common-mode failures. The use of genetic programming is proposed to generate multiple software versions by varying parameters of the genetic programming algorithm. An environment is developed to generate programs for a controller in an aircraft arrestment system. Eighty programs have been developed and tested on 10000 test cases. The experimental data show that failure diversity is achieved, but for the top performing programs its levels are limited