Imitation Learning (IL) to Solve Wordle

OpenAI Gym for Wordle, plus Behavioural Cloning model applied for basic IL.

Code Architecture

the folders checkpoints, results, training_information are for IL:

• checkpoints: stages of the model saved during training at diff epochs (20 cps).
• results: text output of each cp being tested on 100 random games.
• training_information: performance metrics recorded at each cp.

data folder:

• game_history: previous games of wordle used as expert demos (collect.py parses all the info to trajectories_all.npy).
• valid_guesses: words that are valid guesses but not solutions.
• valid_solutions: words that can be guesses AND solutions.
• all_words: combination of both lists (used words_processing.py to combine)

General:

• Wordle.py: code to create gymnasium environment.
• util.py: util functions to create sessions for player and model for one game of wordle.
• main.py: makes a session to let player/model compete with each other.
• test.py: testing model on 10 random games after being trained.

Acknowledgements

• gym-wordle
• Wordle
• Wordle Solving with Deep Reinforcement Learning
• Wo[R]d[L]e