Aura-Farmer

Inspiration

Short Deck Omaha takes poker and moves the furniture. Thirty-six cards, five in your hand, and a flush that outranks a full house. Most of what we knew about poker didn't apply, so we figured a bot had a better shot than we did.

What it does

Aura-Farmer doesn't guess. Before every move, it plays the hand out a few hundred times, works out what its opponent is probably holding, and takes whichever of its four options makes the most money on average. It also watches how each opponent plays and keeps adjusting as the match goes on.

How we built it

Our first version already beat the practice bots the organizers gave us. We didn't trust that. We went through it line by line and found it had been miscounting: it missed its opponent's last move in a lot of hands, so it thought people folded far more than they did. It had been confidently wrong for a while.

After that, we stopped trusting anything we hadn't measured. Every idea had to survive a test on deals the bot had never seen, and most didn't. We threw out more changes than we kept. To test properly, we built ourselves more than twenty opponents, some designed purely to beat Aura-Farmer.

Challenges we ran into

Those opponents found our biggest flaw quickly. Our bot only bet when it had a good hand, so every time it checked, it was basically announcing it was weak. The smarter bots started betting into every check and collecting.

Teaching it to call those bluffs helped. Teaching it to raise them backfired immediately, because the bots we'd built to set traps were waiting. So now it calls. It has stopped trying to be a hero.

Accomplishments that we're proud of

It's steady where it counts. It uses under half its time limit, so a slow tournament machine won't change how it plays, and it hasn't crashed once in hundreds of thousands of practice hands. We're also proud of how many of our own ideas we were willing to kill when the numbers said no.