Inspiration

Poker turns incomplete information into a decision-making challenge. Every bet requires balancing the strength of your cards, the uncertainty of your opponent’s hand, and the cost of being wrong. We wanted to explore how an AI could learn that balance through simulated play.

We built Po for the TAMU Datathon Poker Bot Challenge, where agents compete in heads-up Short Deck Omaha.

What it does

RiverMind evaluates its cards and the betting history to choose when to fold, call, or raise. Its strategy is built around Monte Carlo Counterfactual Regret Minimization (MCCFR), an approach that improves decisions by estimating how alternative actions could have performed.

The challenge uses a 36-card deck, five hole cards, and modified hand rankings. Po must navigate these rules while making decisions within a strict eight-second thinking budget per 1,000-hand half.

How we built it

We used Python and NumPy to build a strategy based on a betting tree and card abstractions. The betting tree represents legal action sequences, while card abstractions group similar situations to make training manageable.

Training produces strategy tables that the bot loads before play. During a match, it identifies the current situation and selects an action from the corresponding strategy.

We explored MCCFR variants that filter low-probability actions and adapt between strategies using observed opponent behavior. We evaluated candidates through reproducible matches and packaged them with the starter’s FastAPI and Uvicorn server for Docker deployment.

Challenges we ran into

The biggest challenge was balancing strategic complexity with runtime limits. Short Deck Omaha creates many possible card combinations, and the bot must make thousands of decisions with limited CPU time.

Card abstraction helped reduce that complexity, but introduced a tradeoff between efficiency and strategic detail. We also had to distinguish meaningful improvements from favorable cards. Duplicate matches, swapped seats, and confidence intervals helped us evaluate results more carefully.

Deployment added another constraint. Everything needed to run on one CPU with limited memory, a read-only filesystem, and no internet access during play.

Accomplishments that we're proud of

Both of our tested MCCFR candidates passed Docker build checks, HTTP smoke tests, and full 2,000-hand validation matches with zero reported errors or forfeits.

In those runs, each candidate used less than one second of charged thinking time per half and stayed below 150 MB of peak memory, leaving substantial room within the tournament limits.

What we learned

We learned how regret minimization turns comparisons between possible actions into a learned strategy. We also saw how strongly an agent’s performance depends on its representation of the game and the opponents used during training.

Testing taught us to consider reliability alongside playing strength. A competitive strategy needs predictable startup, legal actions, and enough time remaining to finish every match.

What's next for Po

We want to refine the card abstractions, train against a broader range of opponents, and improve how the bot adapts during a match. Our goal is to turn those improvements into consistent results across unfamiliar opponents and unseen deals.

Built With

Share this project:

Updates

Submission history