Poker AI

For the first project I created a full-fledged Texas Holdem Poker AI. Perhaps the hardest part of this task was creating the hand evaluator. Poker hands do not follow a ruling system that lends itself easy to numerical assignment. For example, a four of a kind is always better than a full house, by a four of a kind with Aces is much better than a four of a kind with 2s. In order to correct for this non-mathematical heuristic, I created a numerical system that assigned points on the order of the 10,000s for each increasingly better hand (ie. royal flush, straight flush, four of a kind, ...) and then a system that incremented on the order of 10s for same-hand comparison. This way, if two players both got a three of a kind, the three of a kind with a higher face value would be the winner.

Once this hand evaluation class was completed (and painfully debugged), it was simply a matter of creating a game container class that could keep track of the state of the game and record the result. Then, a match making environment was created to handle pairing multiple games of players. At this point, I could play multiple games of poker against multiple human users, but I still needed an AI. While I could have delved much further into the complexity of the AI I created, I settled for one that used 7 key weights to determine what move it would make. To keep things simple, I hardcoded these weights for one AI player and had them randomly generated for the others. For this to be a true evolutionary-learning AI, the next step would be to evolve these weights through generations and tournament style training. However, the way I currently have set up represents a fixed generation of the AI playing against 2 other random generations and winning the majority of games (most of the time).