Update README.md
#1
by
weepingdogel
- opened
README.md
CHANGED
@@ -1,3 +1,73 @@
|
|
1 |
-
---
|
2 |
-
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language: en
|
3 |
+
tags:
|
4 |
+
- pygame
|
5 |
+
- reinforcement-learning
|
6 |
+
- dqn
|
7 |
+
- game-ai
|
8 |
+
license: mit
|
9 |
+
datasets:
|
10 |
+
- custom
|
11 |
+
metrics:
|
12 |
+
- accuracy
|
13 |
+
- reward
|
14 |
+
---
|
15 |
+
|
16 |
+
|
17 |
+
# Big Ball Game AI
|
18 |
+
|
19 |
+
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.
|
20 |
+
|
21 |
+
## Model Description
|
22 |
+
|
23 |
+
- **Input**: Game state vector (69 dimensions) containing:
|
24 |
+
- Player position (x,y)
|
25 |
+
- Player size
|
26 |
+
- Hunger meter
|
27 |
+
- Nearest 13 food items' information (position, size, distance, edibility)
|
28 |
+
|
29 |
+
- **Architecture**:
|
30 |
+
- Dueling DQN with Noisy Linear layers
|
31 |
+
- Feature extraction: 2 fully connected layers (256 units each)
|
32 |
+
- Value stream: 2 noisy linear layers (128 -> 1)
|
33 |
+
- Advantage stream: 2 noisy linear layers (128 -> action_space)
|
34 |
+
|
35 |
+
## Training
|
36 |
+
|
37 |
+
- **Framework**: PyTorch 2.5.1
|
38 |
+
- **Training Data**: Generated through gameplay (~2000 episodes)
|
39 |
+
- **Infrastructure**: CUDA-enabled GPU
|
40 |
+
- **Training Time**: ~4 hours
|
41 |
+
|
42 |
+
### Training Parameters
|
43 |
+
|
44 |
+
```yaml
|
45 |
+
episodes: 2000
|
46 |
+
max_steps: 1500
|
47 |
+
batch_size: 64
|
48 |
+
target_update: 100
|
49 |
+
gamma: 0.99
|
50 |
+
initial_epsilon: 1.0
|
51 |
+
final_epsilon: 0.01
|
52 |
+
```
|
53 |
+
|
54 |
+
## Performance
|
55 |
+
|
56 |
+
The model achieves:
|
57 |
+
|
58 |
+
* Average score: ~3000 points
|
59 |
+
* Win rate: ~40%
|
60 |
+
* Average survival time: 800 steps
|
61 |
+
|
62 |
+
## Limitations
|
63 |
+
|
64 |
+
* May get stuck in local optima (circular patterns)
|
65 |
+
* Performance degrades with very large numbers of food items
|
66 |
+
* Can be overly cautious with larger food items
|
67 |
+
|
68 |
+
## Useage
|
69 |
+
|
70 |
+
Visit [Github Repo](https://github.com/me0w00f/Big-Ball-Swallows-Small-Ball)
|
71 |
+
|
72 |
+
# License
|
73 |
+
This model is released under the MIT License. See the LICENSE file for details.
|