Racing to Perfect AI/ML
If knowledge is power, then having the ability to quickly process and use data is paramount. Even more so is the accuracy and consistency of the data processing. This becomes more critical amidst the rapid promulgation of AI/ML models which rely heavily on data.
The data used in AI/ML models must be evaluated and curated. After all, bad data will yield bad results. For example, if a model is intended to calculate a specific course of action to be taken and the data that is input is corrupted, the likely result would be a course of action taken that could have negative impacts.
Just like humans can learn bad habits and be taught incorrect information, so can an AI model. Therefore, as we move into this high demand for AI/ML, we must ensure our due diligence is performed when acquiring and curating our data.
In a 2021 article in Forbes magazine, Andrew Ng, the founder of DeepLearning AI and co-founder of Coursera, pointed out there needs to be more focus on data quality as the models (code, applications, etc.) have advanced. In other words, focus less on trying to improve the model from a coding perspective and focus more on ensuring quality data is being used to train the models.
Every year in May, my home state holds the Indianapolis 500 and gets me thinking about the speed and stamina that drivers need to have for such a race. Though, it occurred to me that data science has become commonplace in auto racing over the years. As cars are now outfitted with wireless communications systems that transmit data back to the race teams and engineers in real-time, there is an array of data points that must be analyzed quickly to ensure the team and the driver are getting the most out of the car, as well as facilitating decision-making and formulating strategy on the fly.
With sensors providing telemetry data like speed, tire pressure, engine performance, and other critical information, teams can make adjustments to improve the car’s performance and provide the necessary information to the driver to assist with decision-making—like where on the track to push harder, ease up, or simply drive the car better.
Of course, as impressive as all of that is, now add AI and remove the human driver. That’s just what the Abu Dhabi Autonomous Racing League (A2RL) is working to perfect. Their first autonomous race was held in April in Abu Dhabi, without a human driving the Formula 1 cars.
With the aid of cameras, software—and yes, data—eight teams turned their cars to the track. After practice laps and qualifying to allow the cars to learn the track, the four teams that qualified ran an eight-lap race over the course of an hour. While not as quick or as long (distance-wise) as a standard Formula 1 race, it provided a wealth of data for this new kind of racing to advance, as well as showcasing a new frontier for AI.
One of the lessons learned (which can be applied to AI as a whole) is that the AI can only act on the information it is given. This was evident during an on-track spin by the lead car, at which point the trailing cars came to a stop once the caution flag was waived. The rules state that you cannot pass when under caution, so the two trailing cars came to a complete stop. In a typical race, the trailing cars would have slowed under the caution but continued onward, past the leading car that stopped. The order that the cars were running in would have been frozen and, upon restart, drivers would have been re-ordered to match their placement prior to caution. Again, this highlights the cast-iron nature by which the AI is performing based on its programming and the data it ingests.
While this may be quite a niche use for AI/ML and Data Science, it can provide valuable insights to other researchers—those who may be working on autonomous vehicles not confined to a racecourse, for those exploring new use cases and industries seeking to integrate AI, and even for those working to put regulations and guidelines in place to prevent reckless deployment of AI.