PuzzleJAR: Automated Constraint-based Generation of Puzzles of Varying Complexity

Engaging students in practicing a wide range of problems facilitates their learning. However, generating fresh problems that have specific characteristics, such as using a certain set of concepts or being of a given difficulty level, is a tedious task for a teacher. In this paper, we present PuzzleJAR, a system that is based on an iterative constraintbased technique for automatically generating problems. The PuzzleJAR system takes as parameters the problem definition, the complexity function, and domain-specific semantics-preserving transformations. We present an instantiation of our technique with automated generation of Sudoku and Fillomino puzzles, and we are currently extending our technique to generate Python programming problems. Since defining complexities of Sudoku and Fillomino puzzles is still an open research question, we developed our own mechanism to define complexity, using machine learning to generate a function for difficulty from puzzles with already known difficulties. Using this technique, PuzzleJAR generated over 200,000 Sudoku puzzles of different sizes (9x9, 16x16, 25x25) and over 10,000 Fillomino puzzles of sizes ranging from 2x2 to 16x16. ∗Phillips Academy, Andover, MA 01810, USA †Lexington High School, Lexington, MA, 02421, USA 1