| license: mit | |
| language: en | |
| tags: | |
| - reinforcement-learning | |
| - lunar-lander | |
| - tiny-model | |
| - 836-params | |
| pipeline_tag: reinforcement-learning | |
| # LunarLander — 836 params, 100% landings, 0% crashes | |
| A tiny neural network (836 params) that **perfectly solves** LunarLander-v2. | |
| **48× smaller than PPO** (40K params). | |
| ## Performance | |
| | Metric | Value | | |
| |--------|-------| | |
| | Avg reward | **257.7** | | |
| | Full landings | **100/100 (100%)** | | |
| | Crashes | **0/100 (0%)** | | |
| | Params | 836 | | |
| | Training data | 500 episodes (154K samples) | | |
| | Training epochs | 100 | | |
| ## Demo | |
| <video src="https://huggingface.co/Dimitrius174/lunar-lander-836-params/resolve/main/videos/lunar_lander_v2_all.mp4" controls width="600"></video> | |
| ## Architecture | |
| Params: 8×64 + 64 + 64×4 + 4 = **836** | |
| ## Usage | |
| ## Training | |
| Distilled from PPO (stable-baselines3) via supervised learning on 500 expert episodes. | |