CALL US ANYTIME
Joined by Rochester’s leading technology enthusiasts, innovators, business leaders, and more, we had the opportunity to explore the fascinating technology behind reinforcement learning, live in-action with the AWS DeepRacer race car.
Some of the top highlights of the event include:
Featuring three interactive stations — Virtual Reality simulation, AWS DeepLens demo (this camera is used on the DeepRacer and can detect emotional sentiment!), and race car trivia (a crowd favorite) — there was something for everyone!
There’s no denying that TECHTalks Remix: DeepRacer Challenge was one for the books, and we’re grateful for the opportunity to be the first company in the Upstate New York area to bring the AWS DeepRacer for a self-driving race car challenge.
With all the fun stations, you might wonder how the main event- the DeepRacer self-driving race car challenge – came to be! 16 EagleDream employees (and one local student!) made up four teams of developers creating the programming for the DeepRacer to move around the track in the most reliable and speedy manner. Each team created 3 models without ever testing them with the official AWS DeepRacer on a track. All the work was simulated until the flags went down on race time.
1. What was your biggest takeaway from the DeepRacer Challenge?
One of the biggest lessons learned was that how you train a machine learning model is not the straightforward intuitive answer you might think. In fact, it requires a lot of tinkering and iterative design changes. It is really helpful to explore a specific set of variable changes while holding the majority of the other variables constant. I learned having different design philosophies to see what works best in simulated environments was valuable.
Using the agile development methodology complements the iterative nature of model development and iterative sprints. Another key to successful training is to ensure that any simulation environment necessary for training should mimic the eventual real world environment as much as possible. Any variations that aren’t accounted for can cause deviations from the expected outcome.
2. What was your biggest challenge?
The biggest challenge the teams faced was trying to predict success based on a multitude of different factors. How do you weigh the importance of reliability versus speed? How much should you train for the simulation, knowing that the real world will be different? And how do you measure “success” prior to getting on track? Teams tracked their potential models simulated tests in a spreadsheet and found some interesting results.
For the winning team, Coral Caravan, model A had a simulated lap time of around 18 seconds and made it 100% of the way around the track 88% of the time (which was the most reliable. But when it did go off-track, it made it less than 10% of the way around the track. Model B had a simulated lap time of around 12 seconds, only made it 100% of the way around the track 44% of the time, and would go off-track at any number of progress levels. Model C ended up being a more reliable version of Model B that was cloned. Deciding how to make adjustments to these models with different parameters and rewards was the biggest challenge in finding the model that the team believed would perform best on a real track
3. Has this technology from the DeepRacer helped you in working with clients?
Not long after the event, one of the teams actually had a visiting client mention how Artificial Intelligence might be useful in their business. They were able to discuss some of the work each team has done to explore the AI and Machine Learning landscape using the AWS DeepRacer and DeepLens and it really made it clear to the client that our teams are excited and ready to do more in the space. It really helped drive home that our goal is to be a leader and innovator in these emerging technologies and AWS services give us the ability to do so.
We would like to give a huge thanks to our event sponsors — Constellation Brands, Trend Micro, Hoselton Auto Mall, and Amazon Web Services — for making TECHTalks Remix: DeepRacer Challenge possible.
We’re also very grateful for our keynote speaker, Brien Blandford from AWS, for educating us on the technology behind the AWS DeepRacer and all the latest in Reinforcement Learning & Machine Learning!
If you are interested in learning more about DeepRacer, DeepLens, and AWS tools, we would be happy to schedule a consultation with a cloud architect! Schedule a conversation.
Lastly, thank you to everyone who came out to join us! Interested in racing on the track? We are bringing this event back in early 2020 and would love to get more of the community involved on the teams! Contact us here to learn more.