Vibe-coded implementation of the GUBS card game, with RL agent training code. Coded with Claude Sonnet. I give no guarantees that the code is free of bugs, but I did play-test and it seems to work as it should.

GUBS is a card game by Cole and Alex Medeiros (https://gamewright.com/product/GUBS). The aim is to have more 'Gubs' on your field at the end of the game than your opponent(s). It's very fun so do go buy a copy!

gubs_engine contains the engine which sets out the rules for playing a game.

gubs_play allows you to play a game on the command line.

To play against a simple baseline run the code using the following command.

python gubs_play.py --opponent greedy

We also trained a reinforcement learning agent for 1,000,000 episodes. Best performance was around 500,000 episodes, see gubs_analysis/training_log.csv. The weights for this model are included in best.pt. To play against this model run:

 python gubs_play.py --opponent model --weights best.pt

You can train your own model using the gubs_rl script:

python gubs_rl.py train \
    --episodes 1000000 \
    --device cuda \
    --weights-dir gubs_weights \
    --save-every 10000 \
    --eval-every 10000 \

Finally, I was interested in seeing what a good GUBS strategy might look like. The gubs_analysis script contains code to simulate games and collect statistics. A summary output is available in gubs_analysis/summary. The headlines include: the best cards are the Esteemed Elder, Super Lure, and Cricket Song; always interrupt Lures/Super Lures if possible.