Our Aqfer team is full of some pretty impressive humans. Among them is Ross Williams – Aqfer Director of CX Engineering by day, AWS DeepRacer champion by night. 

Ross’s interest in DeepRacer began in 2019, when he decided to enter AWS’s annual competition to deepen his understanding of machine learning. This hobby soon burgeoned into a seriously competitive endeavor. Over the years he improved his model and climbed the DeepRacer leaderboards. Today, Ross is ranked #1 globally in the Virtual Circuit and he won the Global Virtual Champion title in the 2023 DeepRacer virtual championship. This year, Ross has already qualified to revisit the 2024 re:Invent Championship as a 4-time finalist.

What is AWS DeepRacer?

To oversimplify an extremely nuanced topic, AWS DeepRacer is a car racing competition. But we’re not talking soapbox cars, stock cars, F1 cars or Indy cars. 

Designed to help developers hone their machine learning skills, DeepRacer cars are Wi-Fi enabled vehicles that can drive themselves on a track. Developers can participate in physical and virtual competitions to test and improve the models they’ve built to train their self-driving cars.

Reinforcement Learning Used in DeepRacer

DeepRacer uses an advanced machine learning technique called reinforcement learning. In this technique, the learning model reinforces a DeepRacer car for making correct driving decisions such as acceleration and deceleration, turning, etc. It’s kind of how Pavlov trained his dog… if Pavlov was a developer and his dog was a computer making split second decisions.

DeepRacer cars are evaluated on both speed and accuracy. Winners like Ross have built models that combine the best of both to power cars that can accurately navigate a track in the shortest amount of time.

Here’s a quick look at Ross’s champion DeepRacer in action.

3 Lessons in Machine Learning

It’s safe to say Ross has learned a lot about machine learning during his 4-year journey to DeepRacer glory. While he may remain tight-lipped on comprehensive secrets to his success on the track, we convinced him to share some lessons learned in machine learning.

Ross is our resident AI expert, working towards ensuring Aqfer’s Marketing Data Platform maintains its competitive edge in a landscape where AI advances are accelerating at lightning speeds. 

For those of you hoping to make use of big data to supercharge your AI and ML capabilities – ready your pens and take notes. Here are are 3 lessons in machine learning that Ross shared with us, based on his experience with DeepRacer, and applying ML techniques to Aqfer products: 

Develop an Experimentation Mindset

A winning ML model is always a work in progress. Your work as the developer of that technology is never done. While your model is always learning, you should always be tinkering, running small experiments and looking for small optimizations. Whenever an experiment leads to an improvement, focus there – iterate on top of that area until you’re seeing the performance you want. Over time, these experiments will lead to a smarter, more capable product.

 

Quality Begets Quality

Or, garbage in, garbage out. This applies to both the information you feed your models (it must be accurate and up-to-date) and your approach to engineering those models. It behooves the developer to really take the extra time to put high quality effort into your model, as those efforts are compounded over time as the model continues to learn and improve.

One Thing at a Time

To create a winning model, don’t focus on too many things at once. If you don’t isolate variables as you build, you’ll never know which enhancement or experiment  informed the model’s growth in the right direction. One small change can have a snowball effect over the course of training, so it’s important to develop a systematic approach to implementing those changes. 

Machine Learning & Aqfer

Ross’s DeepRacer journey has led to multiple new machine learning interests that help advance our work here at Aqfer. DeepRacer has fostered a deep skillset around other areas of machine learning including Natural Language Processing (NLP)  and Large Language Models (LLM). These models ingest large amounts of data to gather contextual knowledge to make informed decisions. The RAG model or retrieval augmented generation, is a popular application of these machine learning models. These models have a general knowledge of words and concepts, and the contextual relationships between them. Together these models are able to generate very in-depth, accurate answers to queries – this is the technology that powers our internal chat bot, Aquabot. 

From a practical application perspective, these ML models advance many Aqfer use cases, including our ability to query large amounts of data to generate deep insights analytics. We also tap into this technology to better understand user behavior to create in-depth user journeys. 

As AI technology continues to mature, companies want to ask more questions of their data, therefore accessing their data more often, which can incur expensive query costs. Partnering with a platform like Aqfer means a cost-effective way to get more from your data. 

We’ve only scratched the surface of the possibilities for machine learning to advance our capabilities at Aqfer. We’re very proud and lucky to have an expert like Ross on our team, pushing the boundaries of what’s possible with this emerging technology. Interested in hearing more about the new advances in data science that we’re applying to our technology? Reach out – we’d love to chat!