Cybercom's Data Scientist Jouni Luoma is the Nordics DeepRacer Winner and Top 8 Finalist of the DeepRacer Championship 2019. This year everything goes virtual. The 2020 Summit Online Qualifier race in the Spring was the first success for Jouni and he secured his place on the Championship Finals this year too. So, what to expect this year and will Jouni be on the virtual podium? It's going to be a tough race and utterly different than last year. Read Jouni's interview and follow his journey with us.
The competition is only virtual this year – has this changed the situation?
The competition has changed in many ways. Last year the race was on a physical track with physical cars trying to get as fast lap as possible on an empty track. The big challenge then was to succeed with transfer from simulation to real as the machine learning models driving the car were trained on a simulator.
This year the race happens virtually on the simulated environment, and the sim to real is not an issue. Now the challenge is added with different racing styles such as Object avoidance and racing against other competitors on the same track.
How have you been preparing for the competition? Any important learnings from last year?
I have been training new models and developing new rewarding and training strategies for a couple of weeks now. The learnings from the last year are on a relatively general level as the racing style has changed. The object avoidance and head-to-head racing have been available already throughout the year, but I must admit I started with those just a couple of weeks ago.
What's your game strategy? And what're your goals for this race?
The first round of the race is Object avoidance, where competition is to get the fastest lap on the track with obstacles. Running out of the track and hitting obstacles will give penalty seconds. There is some balancing to do between a fast model and a stable model, which avoids all the obstacles. I will try to train a fast model rather than too stable. So, taking a bit risk there as only 4 fastest drivers advance to the next round.
Is there something new on DeepRacer compared to last year?
There are new racing styles. Also, the DeepRacer console on AWS has changed from the last year. And I am doing a training on AWS GPU instances outside the console to speed up the training process.
What are your strengths when it comes to Machine Learning (ML)?
I would like to think that I have a bit more theoretical understanding of the machine learning algorithms behind DeepRacer (Reinforcement learning, Proximal policy optimization) than an average competitor. That said, there are also other aspects to race than the bare ML part. It will be a stiff competition.
What are your other expectations on AWS virtual conference – will you be able to join the events?
I have looked at the schedule, and there are lots of interesting sessions on AI, ML and Data analytics. I will be checking about those.
Jouni will be racing on 2nd December at 4 pm. If you want to share the excitement, join the DeepRacer Twitch channel. The whole Makers family wish good luck for Jouni!