Using AI to play Hearts

For our Comps project, we were assigned the task of creating an AI to play some kind of board/card game under the tutelage of Professor Dave Musicant. We chose to focus on playing Hearts because it poses an interesting set of problems to tackle. Namely, we were interested in the mechanic of 'shooting the moon' and how various AI approaches would tackle such a problem.

We created four different AI agents that employ a variety of strategies which span multiple difficulty levels. Our goals going into the project were to both create agents which could compete against intermediate to advanced level human players and to create agents which would be fun to play against, so some of our design choices weakened the agents in favor of making them a fun opponent.

This was a great project that taught us all not just about AI, but also about working in a team and building large scale applications.

Of our agents, our Strong Heuristic agent was by far the strongest, winning over 50% of games. This agent included a variety of high level hearts strategies and the hard-coded style allowed for the agent to follow a cohesive strategy throughout, well suited to the game of Hearts. Our Weak heuristic agent was the next strongest, followed by our CBR agent and then our Monte-Carlo agent. Although we had hoped for the CBR and Monte-Carlo agents to be stronger, they still provide engaging gameplay because they do not follow an obvious strategy and allow for a high degree of variation in the games against them.

To create our CBR agent, we needed a database of games to draw from. Holger Sindbaek of WorldOfCardGames.com generously provided us with 5000 games of training data played on his website. Having played a multitude of games on his website for research, we can highly recommend it to anyone looking to experience how Hearts is played.

The training data was originally formatted as Json files recording each move of each game. We parsed each Json file, reformatting the data on cards played, the players hands, and scores to create a SQL query which was then pushed using psycopg2 to a SQL table we had prepared. Our final table consists of decision points made by a player who ultimately went on to win the current hand. Each entry records all of the information known to that player including the location of all cards, that player's hand, the current trump suit, and a variety of other metrics.

For our project, we started with an existing implementation of Hearts from GitHub and spent a few weeks modifying it. The original code had a lot of redundancies, with variable and method names being reused in different classes despite having different inputs, outputs, and purposes. This made the code difficult to follow and debug, so I cleaned it up to improve clarity and organization. Making the code more structured not only helped with debugging but also made it much easier to implement graphics.

To create the graphical interface, we used Pygame. First, we set up a display window and added a background. We take the player's hand and render the cards as sprites, which allowed for interactions like clicking and moving them. We organized the sprites into different groups so they could be displayed in different areas of the screen as needed. We also implemented a system for handling the player's turn: when it's their move, the game waits for input, and once they click a card, it disappears from their hand, and the remaining cards shift to adjust for the missing space. Resizing the hand was particularly tricky since every time a card was played, all the remaining cards had to be repositioned dynamically.

Once a card is played, it moves into a separate sprite group representing the current trick. When the trick is complete, the winning player collects the cards, and their score updates based on the points earned. Overall, these graphical features make the game more interactive and provide a clearer way to track the flow of play.