Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,59 @@
|
|
| 1 |
-
---
|
| 2 |
-
|
| 3 |
-
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language: en
|
| 3 |
+
license: apache-2.0
|
| 4 |
+
library_name: jax
|
| 5 |
+
tags:
|
| 6 |
+
- deep-reinforcement-learning
|
| 7 |
+
- neuroevolution
|
| 8 |
+
- tpu
|
| 9 |
+
- autonomous-agents
|
| 10 |
+
- artificial-life
|
| 11 |
+
datasets:
|
| 12 |
+
- synthetic-navigation-sim
|
| 13 |
+
metrics:
|
| 14 |
+
- foraging-efficiency
|
| 15 |
+
- pathfinding-success
|
| 16 |
+
---
|
| 17 |
+
|
| 18 |
+
# DeepRL (Standard Edition)
|
| 19 |
+
|
| 20 |
+
**DeepRL** is a robust, neuroevolutionary navigation agent. Unlike the "Micro" tier, the **Standard Edition** is designed for higher-resolution environments (64x64) and possesses enough neural capacity for active pathfinding and "hunting" behaviors.
|
| 21 |
+
|
| 22 |
+
## 📊 Model Profile
|
| 23 |
+
| Feature | Specification |
|
| 24 |
+
| :--- | :--- |
|
| 25 |
+
| **Model Architecture** | 3-Layer Deep CNN (64-Filter Width) |
|
| 26 |
+
| **Parameters** | ~45,000 (~180 KiB) |
|
| 27 |
+
| **Grid Resolution** | 64x64 Cellular Automata |
|
| 28 |
+
| **Input Channels** | 6 (Life, Food, Lava, 3x Signaling) |
|
| 29 |
+
| **Training Steps** | 48,000+ Generations |
|
| 30 |
+
| **Compute** | 16x Google Cloud TPU v5e (TRC Program) |
|
| 31 |
+
|
| 32 |
+
## 🧬 Key Behavioral Upgrades
|
| 33 |
+
The Standard Edition introduces **Metabolism Decay**. Unlike previous iterations that could "camp" on a single food source, this model must actively navigate to new food patches to maintain its mass.
|
| 34 |
+
|
| 35 |
+
- **Advanced Pathfinding:** Capable of detecting food sources up to 32 pixels away.
|
| 36 |
+
- **Lava kiting:** Uses a wider 64-filter "vision" to weave through complex obstacle patterns.
|
| 37 |
+
- **Mass Regulation:** Emergent behavior allows the agent to stabilize its size relative to available resources.
|
| 38 |
+
|
| 39 |
+
## 🚀 Deployment (Inference)
|
| 40 |
+
This model is optimized for modern consumer CPUs. It is designed to run smoothly at **30-60 FPS** on mid-tier hardware.
|
| 41 |
+
|
| 42 |
+
### Hardware Target: "Desktop Tier"
|
| 43 |
+
- **CPU:** Intel Core i5/i7 (8th Gen+) or AMD Ryzen 5/7
|
| 44 |
+
- **RAM:** 8GB - 16GB
|
| 45 |
+
- **Platform:** Works with any NumPy-compatible environment (Python 3.8+).
|
| 46 |
+
|
| 47 |
+
## 🛠️ Loading the DNA
|
| 48 |
+
The weights are stored as a ragged NumPy object array to preserve the specific convolutional shapes required by the JAX/Keras training pipeline.
|
| 49 |
+
|
| 50 |
+
```python
|
| 51 |
+
import numpy as np
|
| 52 |
+
|
| 53 |
+
# Load the Standard DNA payload (Apache 2.0 Licensed)
|
| 54 |
+
dna = np.load("DeepRL_Final_Gen48k.npy", allow_pickle=True)
|
| 55 |
+
|
| 56 |
+
# Weight Indices:
|
| 57 |
+
# [0, 1] - Conv Block 1 (64 Filters, 3x3)
|
| 58 |
+
# [2, 3] - Conv Block 2 (64 Filters, 1x1)
|
| 59 |
+
# [4, 5] - Decision Output (6 Channels, 1x1)
|