| --- |
| language: en |
| license: apache-2.0 |
| library_name: jax |
| tags: |
| - deep-reinforcement-learning |
| - resnet |
| - multi-agent-systems |
| - tpu |
| - swarm-intelligence |
| datasets: |
| - competitive-foraging-sim |
| metrics: |
| - multi-agent-survival |
| - resnet-efficiency |
| --- |
| |
| # Large-DeepRL (ResNet Edition) |
|
|
| **Large-DeepRL** is a high-capacity, multi-agent reinforcement learning model. It represents the "Predator" tier of the DeepRL evolution series, utilizing a **Residual Network (ResNet)** architecture to navigate a 128x128 high-resolution arena. |
|
|
| ## 📊 Model Profile |
| | Feature | Specification | |
| | :--- | :--- | |
| | **Architecture** | Deep ResNet (Residual Skip Connections) | |
| | **Grid Resolution** | 128x128 (16,384 spatial cells) | |
| | **Parameters** | ~185,000 (~740 KiB) | |
| | **Agents** | 10 Competing Seeds per Environment | |
| | **Input Channels** | 8 (Life, Food, Lava, 5x Signaling/Memory) | |
| | **Training Steps** | Overnight Evolution (Gen 50k+) | |
| | **Compute** | 16x Google Cloud TPU v5e (TRC Program) | |
|
|
| ## 🧬 Architectural Breakthroughs |
| This model moves beyond simple convolutions by implementing **Skip Connections**, allowing the gradient to flow through deeper layers without vanishing. |
|
|
| - **Global Spatial Reasoning:** The 128x128 grid provides 4x the territory of the Standard model, requiring the agent to plan long-distance paths. |
| - **Multi-Agent Competition:** Trained in a "scarcity" environment where 10 agents compete for limited food patches. This forces the emergence of aggressive, high-speed foraging behaviors. |
| - **8-Channel Alignment:** Optimized for TPU HBM alignment, ensuring maximum hardware utilization and zero memory padding bloat. |
|
|
| ## 🚀 Deployment (Inference) |
| While technically runnable on high-end CPUs, this model is specifically targeted for **Low-End GPUs** to maintain real-time performance. |
|
|
| ### Hardware Target: "GPU Tier" |
| - **Minimum GPU:** NVIDIA T4, RTX 3050, or equivalent. |
| - **Alternative:** High-end multi-core CPUs (AMD Ryzen 9 / Intel i9). |
| - **RAM:** 16GB minimum recommended. |
|
|
| ## 🛠️ Loading the DNA |
| The model is saved as a structured NumPy object array. Note the 8-channel input requirement when setting up your inference environment. |
|
|
| ```python |
| import numpy as np |
| |
| # Load the Large-DeepRL DNA (Apache 2.0) |
| dna = np.load("Large-DeepRL.npy", allow_pickle=True) |
| |
| # Architecture Structure: |
| # - Entry Convolution (64 filters) |
| # - ResNet Block 1 (Add + Activation) |
| # - ResNet Block 2 (Add + Activation) |
| # - 1x1 Strategy Head |
| # - 1x1 Decision Output |
