| license: other | |
| license_name: mit-attribution | |
| tags: | |
| - "2048" | |
| - reinforcement-learning | |
| - n-tuple-network | |
| - game-ai | |
| language: | |
| - en | |
| # 2048 N-Tuple Network Model | |
| Trained using TD(0) afterstate learning with 8 six-tuple patterns and 8 symmetries. | |
| ## Stats | |
| - Games trained: 1,200,000 | |
| - Max tile reached: 16384 | |
| - Patterns: 8 six-tuples with 8 symmetry transforms each | |
| - Weight table size: ~347 MB | |
| ## Files | |
| - `weights.bin` - raw Float32 weight tables (8 x 11390625 floats) | |
| - `config.json` - model architecture and training metadata | |
| - `patterns.json` - tuple pattern definitions | |
| ## Usage | |
| Load the binary weights into 8 Float32Arrays of size 11390625 each. | |
| For each board state, compute the feature index for each pattern under all 8 symmetries | |
| and sum the corresponding weight values to get the board evaluation score. | |
| Pick the move whose afterstate has the highest (reward + evaluation). | |