Q-Learning and DQN Reinforcement Learning Project
This project implements tabular Q-learning and Deep Q-Learning using Gymnasium environments.
Task 1: Taxi-v4
Taxi-v4 was chosen because it has a small discrete state space and action space. This makes it suitable for tabular Q-learning.
- Observation space: Discrete(500)
- Action space: Discrete(6)
- Algorithm: Q-learning
Task 2: ALE/Seaquest-v5
Seaquest was chosen because it is a more complex Atari environment with high-dimensional image observations.
- Observation type: image frames
- Algorithm: Deep Q-Learning
- Network: CNN
- Techniques: replay buffer, target network, epsilon-greedy exploration
Results
Taxi learned a useful policy and achieved positive evaluation reward.
Seaquest was trained using a shortened Colab-friendly run. The DQN pipeline works, but longer training would be needed for a stronger Atari policy.
