Jouni did it again! He qualified from the knockout and were the second fasted in his group out of 14 racers. Congratulations! That means that he is in the top 32 finalists in this worldwide Championship, for the qualifications attended over 10000 competitors earlier this year. DeepRacer League is the world's first global autonomous racing league, driven by machine learning.
Jouni Luoma is a Data Scientist and Machine Learning (ML) competence lead in Cybercom Finland. It's for sure that he has a real passion for the ML and is a very skilled Data Maker. Right after the knockouts, we asked Jouni about his feelings and how did the model work.
"Oh yes, I'm satisfied with the result! We had a practice run on Monday, and the model was a bit unstable but relatively fast. I continued the training after that but was not able to improve that much to change the model.
The basic idea of the reward function has stayed more or less the same with all my models. Still, I have played around with the hyperparameters and the speed/steering metadata. The models did not entirely work as I wanted, and I found a place for improvement in my code yesterday but ran out of time trying to train a new model. Luckily the problem was not that large, and the old model still worked OK."
"Box of Doom" on sector 2 caused trouble to many racers, but Jouni managed that perfectly without getting stuck. How did that happen?
"I had to use the manual speed-up multiplier on console for that to happen. In rehearsals I got stuck to that box for 0.5 min or so."
Next round is a head-to-head elimination. Jouni has started to train a model for that for a couple of days ago, just in case. It's a first time for him training a head-to-head model, so no one knows the results, and it’s a mystery for him too.
"It's a speed race, but you cannot hit the other car or drive out. I have not trained a dedicated model for that. I once submitted a time trial version to the race, and it was not competitive. The competition is tough. There are lots of racers who have put numerous hours on training their models throughout the year."
We know, virtual is cool, but like Jouni we all miss the live event excitement: "I miss the atmosphere of the Championships in Las Vegas from the last year. This virtual race is still fun, but not the same! I hope we will have live races again next year."
The race continues 5th December. We wish all the best for Jouni for next round. And we all want to see him in the Grand Prix Finale and again on top 8 finalists!
If you missed Wednesday’s knockout race you can watch Jouni’s 3 minutes performance below (starts 13:25) or on Twitch.tv.
Read more about Cybercom AWS services and solutions.