Teaching AI to Game using RL (TAG)

A research project comparing DQN-family reinforcement learning algorithms trained to play a game from scratch. Three architectures Dueling DQN, Double DQN (DDQN), and Dueling DDQN are implemented, trained, and evaluated, with full statistics and pre-trained models included.

---

Algorithms

Algorithm	Key Idea
Dueling DQN	Splits Q(s,a) into V(s) + A(s,a); better value estimation in dense action spaces
DDQN (Double DQN)	Decouples action selection from Q-value evaluation to reduce overestimation bias
Dueling DDQN	Combines both improvements for more stable, accurate learning

Each algorithm has a Save and Load variant — train from scratch or continue from a checkpoint.

---

Repo Structure

Code/
  ddqn                     - Save.py # Train DDQN from scratch
  ddqn                     - Load.py # Resume DDQN training from checkpoint
  dueling_dqn              - Save.py
  dueling_dqn              - Load.py
  dueling_ddqn             - Save.py
  dueling_ddqn             - Load.py
models/                    # Saved model checkpoints
statistics/                # Training logs and reward curves
Videos/                    # Agent gameplay recordings across training stages
Stats.xlsx                 # Aggregated training statistics for cross-algorithm comparison

---

Running

pip install -r requirements.txt

# Train DDQN from scratch
python "Code/ddqn - Save.py"

# Resume from checkpoint
python "Code/ddqn - Load.py"

Swap ddqn for dueling_dqn or dueling_ddqn to run the other variants.

---

Tech

• Language: Python
• Libraries: TensorFlow / Keras, NumPy, OpenAI Gym
• Tracking: Custom stats logging → Stats.xlsx + matplotlib

---

Context

Built as part of academic research into how different RL architectures learn game playing behaviour from scratch, no prior knowledge, just reward signals and time.