Big Ball Game AI

This model is a Deep Q-Network (DQN) agent trained to play "Big Ball Swallows Small Ball", a dynamic arcade-style game where the goal is to eat smaller balls while avoiding larger ones.

Model Description

Input: Game state vector (69 dimensions) containing:
- Player position (x,y)
- Player size
- Hunger meter
- Nearest 13 food items' information (position, size, distance, edibility)
Architecture:
- Dueling DQN with Noisy Linear layers
- Feature extraction: 2 fully connected layers (256 units each)
- Value stream: 2 noisy linear layers (128 -> 1)
- Advantage stream: 2 noisy linear layers (128 -> action_space)

Training

Framework: PyTorch 2.5.1
Training Data: Generated through gameplay (~2000 episodes)
Infrastructure: CUDA-enabled GPU
Training Time: ~4 hours

Training Parameters

episodes: 2000
max_steps: 1500
batch_size: 64
target_update: 100
gamma: 0.99
initial_epsilon: 1.0
final_epsilon: 0.01

Performance

The model achieves:

Average score: ~3000 points
Win rate: ~40%
Average survival time: 800 steps

Limitations

May get stuck in local optima (circular patterns)
Performance degrades with very large numbers of food items
Can be overly cautious with larger food items

Useage

Visit Github Repo

License

This model is released under the MIT License. See the LICENSE file for details.