OrbitWars

Overview

This project is about building a high-performance reinforcement learning agent for the Kaggle OrbitWars competition. The competition tasks participants with controlling a fleet of ships to conquer a solar system on a 100x100 2D board. Our implementation focuses on leveraging Jax for parallelization of environment steps to accelerate agent training.

Kaggle Competition Mechanics

• Objective: Conquer planets by launching fleets. The player with the most ships at the end of 500 turns wins.
• Environment: A 100x100 continuous space with a central sun (destructive to fleets) and rotating planets.
• Dynamic Elements: Inner planets orbit the sun, and elliptical comets pass through the board, acting as temporary capture points.
• Combat: Continuous collision detection where the largest force on a planet survives the encounter.

Our Approach

• JAX-Native Implementation: We developed a specialized environment wrapper using gymnax and jaxtyping to enable hardware-accelerated training.
• Abstraction & Simulation: The agent shall receive abstracted informations that focus on current and future shipcounts of planets. To this end, we leverage Jax for fast simulation of the environment.

Tech Stack

• Core: Python, JAX, jaxlib
• Simulation: kaggle-environments, gymnax
• Deep Learning: PyTorch