Catan AI

Full Write-Up

Group Members

Tom Hocquet, Ian Zane, Kai Stern

Abstract

In this project, we aim to develop a reinforcement learning (RL) model that can learn to play Settlers of Catan, a multiplayer board game about trading and resource management. Specifically, we aim to use this model to note how the evolution of group dynamics and play styles in a set of agents that repeatedly play each other might be different than agents trained for a more game theory optimal strategy. Catan has a whooping 244,432,188,000 possible starting states, meaning that our agents cannot just memorize games but learn the strategies. We used a virtual environment to train a singular cohort of agents. The same set of agents will repeatedly play against each other, creating an environment similar to how many people would play board games. Below is a gif of the model learning.

Website explaining the rules of catan

Catan Implementation

This was implemented using an adapted version of the Catan AI developed by Karan Vombatkere. The original is accessible at the link below: Karan Vombatkere’s Catan AI

Files

Catan-AI

Folder containing the edited versions of Karan Vombatkere’s Catan-AI plus our modified or implemented files.

code

Folder containing the code used in the project.

• AIGame.py: Implementation to play a game with just the heuristic AI developed by Vombatkere.
• QLearningAIPlayer.py: Our implementation of the agent using Q-learning.
• QLearningGame.py: Implementation to play a game with just Q-learning agents.
• board.py: The class that manages the Catan game board.
• catanGame.py: Implementation to play a game with one heuristic AI and the rest humans.
• gameView.py: Pygame display for the Catan game.
• heuristicAIPlayer.py: Original implementation of the heuristic AI.
• player.py: Base player class which is expanded upon for AI players (either heuristic or Q-learning).