This is group project was completed in Fall 2021 for the Project 2-1 course at my bachelor’s Data Science and Artificial Intelligence at Maastricht University.

Phase 1: Developing the Game

In the first phase of the project, we spent our time developing the game in Java and laying out the backbone for the game-playing agents. We utilized many useful design patterns to make sure the game mechanics, the bot, and the user-interface were mainly disconnected and could be developed simultaneously.

This GIF shows an example of the product at the end of this phase.

Phase 2 and 3:

The remaining two phases were spent both on an improved user-interface and the development of two bot players. For the bots, we used Double Deep Q-Learning in order to estimate the Q-values of certain actions. Since the game is very stochastic, partially observable, and multi-agent, some assumptions had to be made in order to make developing a bot feasible. Instead of being concerned with global actions, the agent learned, for example, the number of troops to choose for an attack. This is still useful for the bot while playing the game, but simplifies the state and action spaces dramatically.

This GIF shows an example of the final product.

If you are interested in the details of the project, the final report for this project is attached below!