On March 3, twenty teams of data scientists and drivers converged on the famed Sonoma Raceway in Northern California – not to see who could be the fastest, but who could drive the furthest on a gallon of gas. You would think that the Toyota Prius has already set a high bar for mileage in the commercially available hybrid car category, but Toyota Research Institute (TRI) wanted to challenge Silicon Valley to make something great even better – by using machine learning. We were thrilled to represent NTT i3 in this competition.
The Same Starting Line for All
Every team was provided with identically configured Prius vehicles, as well as a standard set of tools that included:
- An application for tracking telemetry data to optimize real time race performance.
- A driving simulator that allowed team members to practice driving styles through a virtual program.
Understanding the racetrack and time requirements was another important part of the equation. The challenge was not only to achieve the highest average MPG over 9 laps within a total elapsed time of 40 and 42 minutes (not exceeding 65 MPH), but to do this while simultaneously ‘competing’ with other cars on the track.
From our past work with IndyCar Racing and others, we knew that we had the knowledge and domain understanding to take a data science approach to the challenge – if we had access to the right quantity and quality of data. TRI provided data that they captured during pre-competition laps, but this had some gaps. For this reason, our team decided to take a two-pronged strategy approach, with (1) data science acting as the primary strategy and (2) a practical approach serving to validate our assumptions and models. This would enable us to validate a strategy within the real-world scenario.
- Data Science Approach: We used the existing data sets provided by TRI, and then applied our domain experience and IP to reveal driving insights to inform our strategy for the race course and conditions. We wanted to know how our drivers could use the 4-inputs available to them (ev-mode, throttle, brake, and gear) as effectively as possible on various sections on the racetrack, in order to get the best fuel efficiency. From our past experience, we knew that one of the problems of automotive use cases is in understanding how to use huge datasets that represent both good and bad performance, and then deliver results to the driver as to what they should and should not do on each sector of the track. We used both supervised and unsupervised learning techniques to find these answers.
- Practical Approach: We set aside time to get to know the specifics of the race track and the best driving style for a Prius. In addition to learning the intricacies of the Sonoma race track, we also ran our data science informed strategy through a driving simulator and rented a Prius ahead of the race to test and tweak some of our MPG-optimizing theories.
All of the competing cars lined up at the start line, and when the green flag was waved, we took off – very slowly. After all, this was a best MPG ‘race.’ Our biggest unknown for this competition was the actual traffic on the track and how that would affect our strategy. This was an unknown out of our control in the planning stage. Given our pre-race strategy, it would take some real-time learning to overcome issues of other cars getting in the way at times we needed to go faster, go slower, switch into another mode, or move in and out of the pit stop. There were definitely challenges that negatively impacted our results, but that is also the case in the real world.
The Winner’s Circle
Our data-driven strategy proved to be solid in practice as our final average MPG result was 83 MPG. This was very close to the 86 MPG achieved by the Toyota Dream Team led by the ‘Father of the Prius’, Toyota’s chairman Takeshi Uchiyamada. We were excited to be included in the Winner’s Circle with the ‘Best Machine Learning’ award.
Race for the Near Future
TRI is running the Prius Challenge again next year. There will be an Autonomous Prius that act as the MPG pace car, which the competing teams will need to beat with a human driver. We can’t wait to take our learnings from this year and join this next competition.