Reinforcement Learning for Hangman Game
Designed a Deep Q-Network (DQN)-based agent to play Hangman by predicting optimal letter choices, using RNNs for memory.
RL ML Games Deep Learning
Tech Stack
Python PyTorch Reinforcement Learning DQN RNN
Results
Achieved 85%+ win rate using techniques such as e-greedy strategy, target policy networks, and dynamic learning rate adjustments.
Key Ideas
- Formulated Hangman as a reinforcement learning problem
- Used Deep Q-Networks (DQN) for value function approximation
- Implemented RNNs to maintain memory of previous letter choices
- Applied e-greedy exploration strategy for balanced exploration-exploitation
- Used target policy networks for stable training
- Implemented dynamic learning rate adjustments
Overview
This project applies reinforcement learning to the classic Hangman game, training an agent to make optimal letter choices to maximize its chances of winning.
Problem Formulation
The Hangman game was formulated as a Markov Decision Process:
- State: Current word pattern, guessed letters, remaining attempts
- Action: Choosing the next letter to guess
- Reward: Positive for correct guesses, negative for incorrect guesses, large positive for winning
Architecture
The agent uses a Deep Q-Network (DQN) architecture with:
- RNN component: Maintains memory of previous letter choices and their outcomes
- Q-Network: Estimates the value of each possible action (letter choice)
- Target Network: Provides stable targets for Q-learning updates
Training Techniques
- E-greedy strategy: Balances exploration of new letter choices with exploitation of learned knowledge
- Experience replay: Stores and samples from past experiences for stable learning
- Target policy networks: Separate target network for stable Q-value estimation
- Dynamic learning rate: Adjusts learning rate during training for better convergence
Results
The trained agent achieved an 85%+ win rate, demonstrating effective learning of letter frequency patterns and strategic guessing.