Can we make an RL agent play Mastermind?

Mastermind is a popular code-breaking boardgame for two players which resembles Cows and Bulls.

Contributors :

• Nithish Bolleddula

• Shreejaya Bharathan

Project Summary:

Using Reinforcement learning to train an agent to play the mastermind game. We implemented two algorithms

1. Q-learning
2. Policy Gradient method

Components of Reinforcement Learning

• Environment : Mastermind game board
• Agent : Plays a move guessing the pattern
• Example state : State consits of all the guesses taken by the agent and feedbacks obtained from environment previously.
• Example reward : Tuple of (Number of colors guessed correctly in right position, Number of colors guessed but in wrong position)

We have tried using Q-learning and policy gradient algortihms to make the agent play the game efficiently.

Links to relevant sections:

• Environment for Q-learning
• Q-learning agent
• Training Q-learning agent
• Policy Gradient approach
• Final Presentation

Here is a link to a video we made to explain this project

References:

• An optimal mastermind strategy - https://arxiv.org/pdf/1305.1010.pdf
• https://en.wikipedia.org/wiki/Mastermind_(board_game)
• https://github.com/egeromin/mastermind/
• https://twitter.com/raymondh/status/1250998905984045056?lang=en