Application Study of Collaborative Multi-Model in Wordle Game

This study examines Wordle, a daily word puzzle game from The New York Times, focusing on its features and player behavior through an analysis of its crossword dataset. A neural network model was developed to predict outcome distributions for future dates, estimating the percentages of potential guesses (1, 2, 3, 4, 5, 6, X) using a sigmoid activation function and letter frequency-based weights. The model's effectiveness was validated by comparing average absolute percentage error (MAPE) values across MLPClassifier, RandomForestRegressor, and XGBRegressor, while addressing inherent uncertainties from limited datasets. Additionally, solution words were categorized by difficulty, with a specific assessment of the word “ERNIE,” enhancing our understanding of the factors influencing word difficulty and guiding future guessing strategies to optimize player performance.