Cybercom’s Machine Learning specialist Jouni Luoma won the Nordics DeepRacer League in Stockholm and is one of the DeepRacer finalists in Las Vegas. Jouni has been busy with his studies, but now he has turned his focus to training new models to succeed on the new track. We interviewed Jouni to hear his thoughts just before the re:Invent week. AWS re:Invent 2019 starts 2nd December in Las Vegas, Nevada.
AWS re:Invent 2019 is almost here. Tell us a little about your preparations for the upcoming final?
- Actually, I have been quite busy with my studies and have not had time to make any laps with the physical DeepRacer car. I have mostly competed on virtual league and read more theory of reinforcement learning. To get a nice touch to physical car again, I participated a DeepRacer Hackathon, that was arranged by our customer in Sweden last week. This gave me good ideas on where to focus during last week before the finals. The race track on re:Invent 2019 will be different from the one we used in Stockholm and I cannot use the winning model anymore. Currently I am training new models for the final.
What have you learned within this DeepRacer journey or what are the best takeaways for you?
- I have learned some reinforcement learning, but the biggest takeaways are more related to the infrastructure and how to get these things running on AWS and in local computer when not using DeepRacer console. The DeepRacer console itself was an easy place to start with, but for the competition one needs to tweak more parameters than are available for tweaking in console.
You also participated to the virtual competitions?
- I participated in the Virtual league which consisted of 6 monthly competitions. The virtual league on my part went great even though I did not have that much time to spend on DeepRacer during the last races of the season. The results are not final yet, but on unofficial scores I finished 10th in the league on overall points.
What are your goals to finals?
- The track in re:Invent 2019 is a new one and different from the race in Stockholm. I do not have any information on which kind of times are possible. Also, there will be tough competition. My goal is to not get eliminated on the first round and after that we will see...
What is the key thing to ensure that you are at the podium in the Finals ;) (your strategy if you dare to share – maybe not too detailed info but something!)
- The DeepRacer development team published a paper on sim2real experimentations and I had similar topics on my mind even before reading that paper. Now I just have to implement the training with some data augmentation / image perturbation.
What are your expectations for the AWS re:Invent event?
- I have not been to Las Vegas or AWS re:Invent before and I am not exactly sure what to expect. DeepRacer related events will take at least 2 days there and I will visit some workshops as well. I am participating to AWS DeepRacer Community arranged bootcamp on Wednesday 4th December, and will give a speech on topic: “Minimum math introduction to Machine Learning side of DeepRacer”. And I just try to enjoy the week and not to plan too much. I will take a short vacation after re:Invent to visit some national parks and sights.
Do you know any of your competitors?
- I know quite many competitors as I have been active on AWS DeepRacer community and many of the participants in finals are part of that community also. I am not sure where the toughest competition comes from, but I think the fastest official times on re:Invent 2018 track (the competition track in summits) were made in Japan. Anyhow, the gap is not that large and there are competent developers coming from all over the world.
Tell about the finals schedule and what’s your plans just before the event?
- The first round of finals is held on Tuesday, December 3rd and I will be racing just after noon. There are 4 groups and from each group 4 fastest times go to second round. The second round will be an elimination cup on Wednesday 4th December, and the finals are on Thursday morning on 5th December. I am testing some of the new training strategies with the old track layout (re:Invent 2018, the track we have at office) and hope that I come up with good ideas on training the model on the new track.
Thanks Jouni for the interview. The whole Cybercom Maker family wishes you all the best for the finals!