On Programs that Learn to Write Programs

Currently, it is humans who write the programs that are matching and surpassing human-level performance on challenge after challenge. With time, we expect more programs to be generated automatically by computers. We begin with an illustration of programming by demonstration from the field of text processing. Here, the computer is tasked with writing a program given examples, e.g., extracting and sorting a list of authors from a bibliography. We have a system that uses textual features of examples to improve on the brute force search over programs. For people who already know how to program, we show that using our system can be faster than writing programs. We then present a model of programs that output programs, more generally, and show how learning is a necessary ingredient that is missing in many approaches for generating programs.