A few weeks ago I started to play around with Machine Learning and the Udacity course Intro to Machine Learning. I had to pick up my Python skills, a language I had used before just a few times to make some scripts. Now I am learning the art of data visualization and handling large datasets.
AWS DeepRacer Scholarship Challenge by Udacity started in August and, even though I am still new to this, I signed up for it.
AWS DeepRacer is a virtual little car that you can teach how to race using Reinforcement Learning. AWS has virtual competitions every month and they give away some prizes, mainly AWS credits. There is also a 1/18th model car available for purchasing so you can compete at in-person competitions at AWS events around the world.
Reinforcement Learning is like the hot & cold game. That game where you hide an object, and the other person needs to find it. That person starts moving randomly without knowing absolutly anything about the location of the object, and you give them hints in the form of «hot» and «cold» depending on how close the are getting to it. After a while, the person looking for the object will start narrowing down the area where it is, eventually finding it.
This is the same. The car starts moving erratically around the virtual world. At the same time, our code will give it «rewards» (like «hot» and «cold») depending on how close it is getting to the goal we set. The thing is, at the beginning the car knows absolutely nothing, it does not know that it should go on the track, it does not know that the goal is to accomplish a full lap, it doesn not even know how to steer… steering? what is that? We let the algorithm to figure all that out. The video below is at the beginning of training.
And this second video after 2 hours and 45 minutes. You can definitely tell the difference.