AWS DeepRacer event for EPAM

EPAM Systems, Inc. is an AWS Gold Partner and has a lot of successful AWS customer launches: https://partners.amazonaws.com/partners/001E000000Rp5QNIAZ/EPAM%20Systems,%20Inc.

A major AWS DeepRacer event organized by Amazon Web Services and EPAM took place recently for the first time. So far, it is the biggest and the most ambitious educational Event EPAM Cloud Practice has conducted in partnership with AWS.

What is AWS DeepRacer?

AWS DeepRacer is a reinforcement learning (RL)-enabled autonomous 1/18th-scale vehicle supported by services in the AWS Machine Learning ecosystem. It offers an interactive learning system for users of all levels to acquire and refine their skill set in machine learning in general and reinforcement learning in particular.

I would recommend going through this link and watching the following video to get a high-level overview of AWS DeepRacer.

In essence, AWS DeepRacer allows a developer to train and deploy a reinforcement learning model onto a small-sized race car (virtual or physical). The performance of the model is measured by the amount of time that it takes for this car to complete from 3 to 5 laps of a simulated (for a virtual car) or real (for a physical car) race track.

A developer can submit his model to the AWS DeepRacer Virtual Circuit contest. The model will be ranked and compared to models from other developers across the globe by the best (smallest) lap completion time of a specific race track. The best racers have a chance to win different prizes including a trip to an offline final race event.

About the event

The AWS DeepRacer event was held in EPAM from September 28 to October 12, 2021. About 1000 EPAMers of all seniority levels from 32 countries joined the competition.

The competition rules are given below:

• Every EPAMer is allowed to join, experience in ML does not matter
• Every participant gets a temporary AWS account for the competition and has 14 hours for AWS DeepRacer service usage
• Every participant can request additional hours if needed
• Top-10 racers from the qualification race reach the final race on a secret track with the model submitted for the qualification.

The rules of the qualification race are in the picture given below.

Results

The qualification race was very competitive because a lot of participants tried to do their best to make it to the finals. Once anyone hit the top-10, they could not feel secure because the list of participants was changing every day! Just imagine, the last changes in the top-10 had happened in less than 1 hour before the qualification race was closed.

According to the event statistics, the participants submitted 3444 training jobs with a total duration of 240923 minutes (~6 months)!

I managed to take 9th place in the qualification race. The leaderboard is in the picture below:

The downside of such a hard competition was that some participants overfitted their models to pass the qualification. To be honest, there were objectively better models with the good ability to generalize and complete laps on different tracks, but they did not reach the top-10 just because they were not fast enough on the qualification track.

The final race was held on the “London Loop Training” track. The final results are in the pictures given below. Unexpectedly, I came on the 4th!

In the next post, I will describe my approach to the event and struggling for the top-10.