Does AI have a methodology which is different from software engineering?

摘要

This paper argues that the conventional methodology of software engineering is inappropriate to AI, but that the failure of many in AI to see this is producing a Kuhnian paradigm ‘crisis’. The key point is that classic software engineering methodology (which we call SPIV: Specify-Prove-Implement-Verify) requires that the problem be capable of being circumscribed or surveyed in a way that it is not, for areas of AI, like natural language processing. In addition, it also requires that a program be open to formal proof of correctness. We contrast this methodology with a weaker form complete Specification And Testability (SAT — where the last term is used in a strong sense: every execution of the program gives decidably correct/incorrect results) which captures both the essence of SPIV and the key assumptions in practical software engineering. We argue that failure to recognize the inability to apply the SAT methodology to areas of AI has prevented development of a disciplined methodology (which is unique to AI and which we call RUDE: Run-Understand-Debug-Edit) that will accommodate the peculiarities of AI and also yield robust, reliable, comprehensible, and hence maintainable AI software.