What can automatic programming learn from theoretical computer science?

This paper considers two (seemingly) radically different perspectives on the construction of software. On one hand, search-based heuristics such as genetic programming. On the other hand, the theories of programming which underpin mathematical program analysis and formal methods. The main part of the paper surveys possible links between these perspectives. In particular the contrast between inductive and deductive approaches to software construction are studied, and various suggestions are made as to how randomized search heuristics can be combined with formal approaches to software construction without compromising the rigorous provability of the results. The aim of the ideas proposed is to improve the efficiency, effectiveness and safety of search-based automatic programming.