Marble Racer - MLAgents, Neural Networks, Unity

This project involves harnessing the power of Unity ML-Agents to train a sphere – my marble racer – to navigate through increasingly complex tracks. I'm tackling this challenge step by step, using transfer learning to progressively introduce more difficult elements into the track, like steeper inclines and tighter curves, to elevate my agent's skills gradually. The marble is not controlled by me or any person, it is controlled by a Neural Network using Unity's ML-Agents framework.


I started with the basics: getting the marble to follow a simple, straight path. Then, like leveling up in a game, I added more challenges to the mix. The goal? To have my racer conquer tracks that throw curveballs every step of the way.


In this training regimen, rewards and penalties are my tools for teaching the racer the dos and don'ts. I give it a virtual pat on the back for hitting checkpoints and crossing the finish line, and I'm not afraid to be strict when it slips off the track. It's all about finding that sweet spot in the reward system to keep the racer on the right path.


As I've progressed, I've kept a close eye on the learning curve, diving into TensorBoard to track our performance. I've been fine-tuning everything: the agent's neural network, the sensors – the whole nine yards – to sharpen its racing prowess.


Once I got a handle on the simpler tracks, I began exploring transfer learning, taking the lessons learned and applying them to more complex circuits. This isn't just about building a smart racer; it's about crafting a versatile one that can adapt to any new track I dream up.


I'm pushing the envelope, building from the ground up to create not just an intelligent agent but a clever, strategic racer that can handle any twist, turn, or loop I throw at it. It's been a journey of iteration, analytics, and a whole lot of trial and error, but the results? They're starting to speak for themselves.