A few tools for playing the most challenging of the daily puzzles on Puzzmo.
- Fully vectorized implementation of Pile-Up Poker (Pro), which enables generating ~100k positions per second for efficient RL training
- Tuned PPO training scripts and pretrained models for Pile-Up Poker and Pile-Up Poker Pro in
pileup_poker/ - Solvers for SpellTower, Wordbind, and Cube Clear in
word_games/
The two included checkpoints, pileup_poker/best4x4.pt and pileup_poker/best5x5.pt were both trained over a period of about 12 hours, and play quite solidly (better than me, anyway) but don't perform any search, so their value estimates and decisions very close to the end of games can be a bit off.
| Model | Average Score / Game |
|---|---|
best4x4.pt |
~$5400 |
best5x5.pt |
~$9300 |
Users can enter each hand as a space delimited case insensitive string of card names (e.g. AS, JH, 10D, 3C), where the wild card joker in Pile-Up Poker Pro is represented as WW .
~/puzzmo_tools/pileup_poker$ python play.py
Play 4x4 or 5x5? [4, 5]: 4
Enter 5 cards for next hand: jc js 9h kd 8s
...
Board:
=====================
| 9H | JC | | 8S |
=====================
| | | | |
=====================
| | JS | | |
=====================
| | | | |
=====================
Hand:
===========================
| | | | | |
===========================
Discard:
=====================
| KD | | | |
=====================
...
The SpellTower / Cube Clear solver runs DFS from every starting point to find all possible words at a given state, then uses length as the search order for a BFS to find a full-clear solution. Game dynamics, e.g. tile clearing and gravity, are factored in end-to-end (except for points / bonus point calculations).
Input grids can be edited in spelltower.py and cube_clear.py. Inputs are entered at runtime for wordbind.py.
All word games use Trie Search on the Scrabble US dictionary, which may or may not exactly match the internal dictionary on Puzzmo.