[Part 1 of 2]
Remember playing with remote control cars as a child, chasing them, mastering turns at speed and occasionally righting it after accidentally flipping it onto its back?
Perhaps you still enjoy playing with remote control cars, but your desire to learn machine learning is just getting in the way?
Well, AWS has stepped up its game to solve all - or at least some - of your problems!
Introducing the AWS DeepRacer, a model car that uses your machine learning models to race around a track.
AWS unveiled the DeepRacer at re:Invent 2018, so it isn’t a new service by any means. If you’re based in the UK however, getting hold of one of the cars has been somewhat challenging. Never fear - you don’t actually need to have your hands on one of the cars to get started with DeepRacer (just an AWS account and a desire to learn Reinforcement Learning).
Reinforcement Learning is a branch of machine learning that focuses on rewarding correct actions, much like if you were training a puppy. If you ask your puppy to sit and it sits, you give it a treat. If it sits on the first try, you might give your puppy lots of treats to really reinforce the behaviour.
DeepRacer is a similar concept only instead of treats, you want the highest possible score.
The aim of DeepRacer is to get your car around the track successfully in the fastest possible time.
A simple reward model to achieve this might be:
10 points - for staying on the track
0 points - for leaving the track
But what about trying to correct behaviour before it strays from the track?
You could split the track into 3 sections and reward the car for staying as close to the centre line as possible. The further from the centre line the car strays, the fewer points earned. This is the example AWS provides you with when you set up your first model.
As with everything Amazon provides, you have control over a whole host of the cars functions.
You can set the maximum angle the wheels can turn, how fast the car can go and even the granularity of speed per wheel angle. As for rewards, there are a number of parameters you can reward such as speed, staying on the track and bearings.
A little overwhelming isn’t it? For me that’s part of the fun! There is so much to experiment with. So many options to try and refine that model and earn the top spot in the DeepRacer League!
Stay tuned for part 2 where we train some models and test them on the virtual track!
To learn more about AWS DeepRacer visit:
Author - Michelle Chismon