Evolving Romanian Crossword Puzzles with Deep Learning and Heuristic Search
Vadim Bulitko, Adi Botea · 2021 IEEE Conference on Games (CoG) · 2021
Crossword puzzles are a challenging game of skill in which humans have been competing for decades. Recently a heuristic-search-based Artificial Intelligence (AI) solver for Romanian crossword puzzles achieved competition-level scores. In this work in progress we tackle procedural content generation of crossword puzzles. Using genetic algorithms we evolve crossword puzzle instances on which the AI solver can achieve a high score. Since the solver takes a substantial time to solve each instance we first train a deep neural network to predict the solution score the solver would achieve. We then run the evolution of crossword puzzles with the network as the fast proxy fitness function. We show that doing so is an effective way of procedurally generating numerous crossword puzzles.