Master's Thesis presentation, Stanislaw Malec

Abstract: This thesis investigates the use of deep reinforcement learning for assessing the difficulty of levels to the match-3 game. Proximal policy optimisation agents with action masking is used for the task. We develop a performant, feature-rich match-3 simulator and conduct experiments demonstrating significantly better performance over a random policy on both seen and unseen levels. Furthermore, our results show that training on a subset with complex levels gives the best generalization performance.