How Machine Learning Engineers Make Puzzle Games More Fun

Devsisters is utilizing a Reinforcement Learning-based PuzzleBot for level balancing in Cookie Run: Witch's Castle.

The existing stage Polishing process tended to rely on workers' intuition, which could be subjective, and had the limitation of becoming less efficient as the number of stages increased. To address this, Devsisters built a stage evaluation metric that can objectively evaluate stages based on PuzzleBot data.

The core criterion of the evaluation metric is the rate at which users continue playing or purchase hearts to retry after failing to clear the game. This represents an 'attractive stage' that users want to clear even if it means paying.

Analysis of data from the previous title, Cookie Run: Puzzle World, confirmed a correlation showing that the fewer remaining missions upon failure, the higher the user's retry rate. Based on this, two types of metrics were designed.

• Overall distribution of remaining missions: A normalized metric that assigns higher weight the closer the remaining mission count is to 0
• Metric from a specific ratio or specific count perspective: The proportion of plays that failed with only missions below a certain ratio of the total mission count remaining