Date: 2020.09
Role: Sole Developer
As the sole developer, the project was developed entirely by myself from design to programming.
About
ML-Shooter is a small-sized experiment on training a game-playing shooter bot using reinforcement learning. Using Unity’s ML-Agents toolkit, I trained a shooter bot that successfully performs the difficult task of aiming and shooting a moving target.
Curriculum Learning
I initially tried training the agent in the environment that I designed as a whole, but the agent wasn’t able to show good performance due to the high difficulty of the task.
So, I decided to use Curriculum Learning, a training strategy that gradually increases the task to effectively train a machine learning model, by introducing more constraints over time, such as moving the target around more complexly, adding ammo penalties, etc., as the training progresses.
Through a carefully crafted curriculum that I created over lots of trial and error, I was able to successfully train the agent to successfully shoot a speed-varying target using limited ammo.