Infoscience

Thesis

Interactive Programming by Example

As of today, programming has never been so accessible. Yet, it remains a challenge for end-users: students, non-technical employees, experts in their domains outside of computer science, and so on. With its forecast potential for solving problems by only observing inputs and outputs, programming-by-example was supposed to alleviate complex tasks requiring programming for end-users. The initial ideas of macro-based editors paved the way to subsequent practical solutions, such as spreadsheet transformations from examples. Finding the right program is the core of the programming-by-example systems. However, users find it difficult to trust such generated programs. In this thesis, we contribute to proving that some forms of interaction alleviate, by having users provide examples, the problem of finding correct and reliable programs. We first report on two experiments that enable us to conjecture what kind of interaction brings benefits to programming-by-example. First, we present a new kind of game engine, Pong Designer. In this game engine, by using their finger, users program rules on the fly, by modifying the game state. We analyze its potential, and its eventual downsides that have probably prevented its wide adoption. Second, we present StriSynth, an interactive command-line tool that uses programming-by-example to transform string and collections. The resulting programs can also rename or otherwise manage files. We obtained the result that confirms that many users preferred StriSynth over usual programming languages, but would appreciate to have both. We then report on two new exciting experiments with verified results, using two forms of interaction truly benefiting programming-by-example. Third, on top of a programmingby- example-based engine for extracting structured data out of text files, in this thesis we study two interaction models implemented in a tool named FlashProg: a view of the program with notification about ambiguities, and the asking of clarification questions. In this thesis, we prove that these two interaction models enable users to perform tasks with less errors and to be more confident with the results. Last, for learning recursive tree-to-string functions (e.g., pretty-printers), in this thesis we prove that questioning breaks down the learning complexity from a cubic to a linear number of questions, in practice making programming-by-example even more accessible than regular programming. The implementation, named Prosy, could be easily added to integrated development environments.

Related material