Computer AI players have already surpassed human opponents in competitive Scrabble, however, defeating a Computer AI opponent is complex and demands efficient heuristics. The primary objective of this project is to build two intelligent AI players from scr atch for the Scrabble cross - board puzzle game having different move generation heuristics and endgame strategies to evaluate their performance based on various benchmarks like winning criteria, quality of moves, and time consumption. The first AI selected is the most popular Scrabble AI, Maven. It generates a three - ply look - ahead simulation to evaluate the most promising candidate move and uses four different heuristics for the fast move-generation. The second AI, Quackle, is the strongest alternative Scrabble AI to Maven. It generates its best candidate move by using a three-ply look-ahead simulation and win probability estimation. In this project, we primarily focus on the end - game heuristics because end-game sessions are complex, real-world situations where the move options are limited and require expert techniques and model strategies to maximize the reward. Moreover, the basic game heuristics used in the mid-game are not sufficient for an end-game. For this project, we created four variants of Maven AI. After conducting experiments, we observed that Maven Q - Sticking slow - end game AI performs better than other AI variants like Maven Q - Sticking AI, Maven slow - end game AI, No Q - Sticking AI and Quackle AI.
Abraham, Priyatha Joji, "A Scrabble Artificial Intelligence Game" (2017). Master's Projects. 576.