Planet Wars is a real-time strategy game where players compete with each other to conquer and defend planets by deploying fleets of spaceships in a dynamic map. The game requires long-term strategic decisions such as when to expand, stay in defense, or attack. In this work, we evolve a neural network controller for Planet Wars using CMA-ES. The neural controller takes a representation of the current game state and outputs decisions such as which planet to target and how many ships to deploy. CMA-ES optimizes the network’s weights using a simple fitness function of win-ratio over 100 games. The objective of this project is to create a competitive agent that can outperform baseline agents via evolutionary computation.

You can access the GitHub repository of our code via this link.