| --- |
| tags: |
| - estimation |
| - card-game |
| - reinforcement-learning |
| --- |
| |
| # kooz-bot nets |
|
|
| Trained nets for **kooz-bot**, a bot for Estimation (the 4-player trick-taking card game) entered in |
| [estimation-arena](https://github.com/essamgouda97/estimation-arena). |
|
|
| `final.pt` is an estimation-engine checkpoint (PyTorch, weights only) holding three nets: |
| BiddingNet (bids, With, estimates, Dash Calls), CardPlayNet (card play) and GuessNet (who holds each |
| unseen card). Trained from scratch, no human games: 1,000 recorded PIMC games for a card-play warm |
| start, then 3,000 loops of self-play against itself, past versions, a rule-based bot and a light |
| PIMC (encoder v3), about 10.6 hours on one laptop. |
|
|
| kooz-bot plays these nets through estimation-engine's Hybrid search: 64 imagined deals per card, |
| placed by GuessNet and kept by the bidding, every legal card searched, and a Dash Call only when |
| BiddingNet values it at +10 points or more. |
|
|
| sha256 of `final.pt`: `d4e104148a0a8fe74e56e6c82670fbee8aade0e9163b122316dc4aeab808fda0` |
|
|
