# AWS DeepRacer ## Team Summary (Provided by ChatGPT) Introducing the Three Not Out team - a group of passionate and driven individuals who are on a mission to build the ultimate DeepRacer model. With a diverse range of backgrounds and expertise, this team brings a unique and innovative approach to the table, tackling challenges with creativity and collaboration. Their dedication to the project is unmatched, and they are constantly pushing themselves to new heights in pursuit of their goal. With their eyes set firmly on victory, the Three Not Out team is ready to take on any challenge that comes their way and prove themselves as true champions. ## Links Login: https://dashboard.eventengine.run/login Event Code: aa81-153882f744-20 ## Individual Notes ### James S * Planning on building a basic reward function considering steering angle to next waypoint, some speed and progress params. Then being disappointed when a 12 year old off the Internet beats me. Didn't do this, working on variants of upday ### Mikey G * https://github.com/hamham1004/DeepRacer - Mikey G * on eval, times were [12.4, 12.4, 10.8] * First 2 laps include a 2 second time penalty * If we can eliminate those, not a bad start * Have cloned and am re-training on a different track * https://github.com/sasasavic82/deepracer-reward - Mikey * https://github.com/rajat5ranjan/AWS-DeepRacer - Mikey ### James B * Trying out some online models - will take what looks like the best option and try out tweaking parameters, training on multiple tracks, etc. * Things to try: * Discrete action space (force a fast speed for straight ahead) * Training on different tracks ### Rob H * Getting a baseline time from the example code * Researching tracking to best line around circuit - https://blog.gofynd.com/how-we-broke-into-the-top-1-of-the-aws-deepracer-virtual-circuit-573ba46c275 - looks like this needs too many hours of training * Tweaking the standard models to try and get some easy wins. ## Questions * Can we automate action space and hyperparameter sweeps e.g. using grid search? * Is a "unbalanced" action space better (i.e. more options to turn left thn right, since there are no big right turns on the track) * Is it better to add or multiply sub rewards? * Will AWS people change things or is it a simple time trial around a given track (in a given direction)? * How can we tell if a model is performing well? * How should we store results and code? * How should we pick which to submit for the race? ## Models Being Tested * https://github.com/MatthewSuntup/DeepRacer/blob/master/reward/reward_final.py - James B * Continuous action space * Discrete action space with extreme right turns removed * My own (!) - James S * Initial Example as baseline on re:Invent2018 track, trained for 1hr, average laptime 22s - Rob * Upday https://medium.com/axel-springer-tech/how-to-win-aws-deepracer-ce15454f594a - James S * Tweaked -15/30 steering range, 4m/s max speed * Tweaked -15/30 steering range, 2m/s max speed * Reward function suggested by ChatGPT - James B * Hard coded racing line - James B | Model Name | Laptime - re:Invent2018 | 3 Laptime - re:Invent 2018 | | ----------- | ---------------------- | --- | | Default Example | 00:22 | | gh-1 (James) | 00:19.5 | 00:59.061 | mg-1 (Mikey) | 00:15.1 | 00:45.331 | ChatGPT (James) | 00:12.863 | 00:46.782 # Important Stuff * re:Invent2018 track * Unlimited virtual track attempts * https://docs.aws.amazon.com/deepracer/latest/developerguide/deepracer-reward-function-input.html * Go round anticlockwise * Need to sign up for slot to test on real track * 1200 Fri - final team model uploaded * Max 4 models in parallel # Resources * https://docs.aws.amazon.com/deepracer/latest/developerguide/deepracer-get-started-training-model.html - https://us-east-1.console.aws.amazon.com/deepracer/home?region=us-east-1#getStarted