PTCG MuZero Agent
MuZero agent for the Pokémon Trading Card Game (PTCG), trained via hybrid
ISMCTS (belief sampling) + collapsed chance nodes for stochastic
transitions.
Architecture
| Component |
Details |
| Representation |
Transformer encoder (4 layers, d=256) |
| Prediction |
Two-head MLP (policy + value) |
| Dynamics |
Residual MLP with stochastic branch |
| Card embedding |
Learned (64d) + static CSV features (48d) |
| Search |
Gumbel MuZero + ISMCTS (4 determinisations) |
| Interpretability |
5 linear probing classifiers |
Training
- Framework: JAX / Flax
- Devices: 2× GPU (data-parallel via jax.pmap)
- Step: 160,000
- Batch size: 64
- LR: 0.0003
Files
| File |
Content |
muzero.safetensors |
MuZero network weights (h + f + g + probes) |
deck_builder.safetensors |
Deck builder policy weights |
config.json |
Full training configuration |
