Inspiration
This Poker Bot Challenge allowed us to test our decision making skills through this program. Since both poker bots play the same deals with their seats swapped, getting lucky can only take you so far. We wanted to see how well our bot’s strategy would perform against others across thousands of hands.
What it does
Our bot, 3/7, plays heads-up short-deck Omaha. It simulates possible opponent hands, community cards to estimate its winning probability, and considers what action is worth taking depending on the price (whether to call, check, fold, or raise). The same hand that is worth playing against a small bet may be worth folding against a pot-sized raise.
How we built it
We used the tournament’s Python starter code and wrote the decision logic in strategy.py. The bot removes every visible card from a 36-card deck, samples five cards for its opponent, and fills in the rest of the board. The evaluator scores each showdown using exactly two private cards and three community cards. We count wins and split pots, then compare our estimated share with the cost of staying in the hand.
Challenges we ran into
The judge engine gives us eight seconds of thinking time for an entire 1,000-hand half. We had to adapt to different types of strategies, and be able to decide what plays were worth continuing in the simulation. For example, if an opponent raises again before a new card appears, the price changes, but the cards are the same. In this case, we would need to reuse the previous estimate and recalculate whether the call is worth it.
Accomplishments that we're proud of
In a 2,000-hand match through the judge, 3/7 finished with zero errors and earned an average of 6.835 big blinds per paired deal. It used 3.73 seconds in the first half and 4.18 in the second, well inside the eight-second limit. We also tested against the different strategy bots that always call or always raise the pot. Our strategy finished ahead of both across all three test seeds.
What we learned
A hand’s chance of winning is only one part of the decision making process. If calling costs 10 chips and brings the pot to 40, winning a quarter of that pot on average is enough to break even on the call. We were also introduced to the limits of our approach during testing. We estimate our chances against a random hand, but an opponent who keeps raising may be telling us something about the cards they hold.
What's next for Poker Bot Challenge - 3/7
Our next experiment is to give stronger opponent hands more weight after repeated raises. We plan to test that change against the current version on the same deals to see whether it helps us fold expensive losing hands without giving up too often.