RL play Worlde

Using RL to train a bot play wordle

Overview

I recently done a project where I used reinforcement learning (RL) to solve Wordle, a popular word puzzle game. By leveraging the power of machine learning, I trained an RL agent to learn how to guess words based on feedback. But this project is not complete yet, there are spaces for improvement and my model can’t run on the full dataset yet.

Project Structure

Training

Training was conducted using A2C. I trained the agent using the first 10 words from the official Wordle vocabulary first and using 10 action to represent the word the model can choose, gradually teaching the model to make better word guesses. After the mean reward is great, I transfer learning to train on first 100 words and 100 action. And after that I train on the full dataset which include 2315 words and 2315 words

Challenges and My Propose Solutions

1. Challenges

I face with a problem is that the training is too long. When I train on the first 10 words. The learning is fast. But when I transfer learning, the model is learning too slow. I spend 7 hours on training 100 words model but it get 25% right mean answer. But it is better than random guess which is 6%.

2. My Propose Solutions


Learn More

For a detailed exploration of the code, and methods, you can view this Kaggle Notebook.