How Yandex.Taxi Algorithms Steer Drivers To Higher Earnings
A couple months ago, Yandex.Taxi debuted its new proprietary “Pathfinder” system that tells drivers exactly where to go in between trips to maximize the chances of getting another ride request fast. Pathfinder predicts demand city-wide and ensures enough cars are always headed in the right direction to meet it. Drivers who follow its suggestions throughout the day boost their earnings by an average of 20%! Lev Feofanov, Director of Experimental Products, wrote an article on why this new “route suggestion system” matters.
Higher Efficiency = Higher Earnings
Yandex.Taxi’s users fall into two ride-hailing industry-standard categories: drivers and passengers. Both of these groups care about how safe, fast and accessible the service is; the biggest difference is that drivers depend on Yandex.Taxi to make a living. That means they want to spend as much time as possible actually driving passengers (i.e. earning money) while avoiding dead mileage.
But today, even in big cities, the average driver only spends 2/3 of their time on the clock with passengers in the car. The other 1/3 is wasted because they’re waiting for their next ride request or driving to a pickup point.
Pathfinder optimizes how drivers spend their time behind the wheel without adding more hours by increasing how often they have passengers in the car.
It’s Hard for Drivers to Make Predictions
Yandex.Taxi is built on a set of algorithms that help drivers maximize efficiency behind the wheel. For example, our dispatch algorithm lets drivers accept another nearby ride request before dropping off their current passenger.
But unfortunately, these types of “stacked rides” aren’t always available, especially if the ride ends in an isolated area where there isn’t much taxi activity.
This is where driver know-how comes into play: should I stay and wait for another ride request here, or drive somewhere else hoping I’ll find one there?
Option 1, waiting around for another passenger, comes with no guarantees. The driver could end up waiting 5 minutes or 50. Option 2, drive somewhere else, doesn’t come with any guarantees either, and there’s a good chance the driver will waste time and gas for nothing.
Drivers are forced to rely on their own subjective knowledge of the city. Over time they learn where the morning, lunchtime and post-work hot spots are, but this is all they have to rely on when deciding where they might be needed next.
To make drivers’ lives easier, Yandex.Taxi uses objective data to help them plan their next move. A few years ago, we started highlighting zones on the driver app map where demand was highest and driver density lowest (surge pricing zones). But this still fell short of solving the bigger “where exactly should I go now?” problem. Drivers often find themselves close by multiple surge zones, but there’s heavy traffic to get there. So where exactly should they go to waste the least amount of gas and time getting there?
In the real world, this decision is so difficult that drivers most often choose to just stay put where they dropped off their last passenger, even if there’s surge pricing nearby.
Helping Drivers Make the Best Decisions
This is where Pathfinder, our newest algorithm, comes flying in to the rescue! It puts an end to the guessing game by setting the optimal route to the closest zone. Statistically speaking, these suggestions are guaranteed to increase daily earnings.
We taught Pathfinder to work by building a virtual city and integrating all our dispatch and pricing algorithms in it. Then we brought it to life by setting loose thousands of virtual people who went about their daily errands just like in real life.
Using machine learning, we were able to model the typical behavior of a driver in this context using their subjective know-how in combination with Yandex.Taxi driver app map surge zones. The more days our virtual drivers worked and learned about the natural pulses of the city, the better their decision-making became and the more they started earning.
Then we unleashed Super Drivers in the city, or virtual drivers who know literally everything about where the best ride requests are, when and where there will be surge pricing, when it ends, and how to reach those zones fastest. In most cases, Super Drivers earned more than regular drivers in the same amount of time on their shift.
We repeated this simulation is various cities in different seasons and amassed a huge amount of data showing a tangible difference between the earnings of regular and Super Drivers who know everything before it happens and stick to strategically optimal routes.
Then we just took decision making system and reworked them into an optimal route recommendation system (Pathfinder) that now helps regular drivers work just as efficiently as their omnipotent virtual colleagues.
How Our Driver Distribution Algorithm Works
We break down the city into thousands of tiny zones, see how many users are in each of them, crunch past use statistics, and consider data on road conditions and upcoming events (football, concerts, etc.). This forms the base we need to estimate how many cars will be needed where to meet demand.
Once we estimate demand over the next few hours, the next step is rounding up nearby drivers and suggesting they take the optimal route to where we predict they’ll get ride requests.
For example, we know that after 7:00 p.m. there will be about 150 passengers looking for rides in the south of the city. We see there are maybe 70 drivers dropping off passengers nearby, so we need to find another 80 to cover demand and prevent surge pricing.
This is where things get tricky. We need to consider several unpredictable factors at once: 1) drivers aren’t required to follow our recommendations (if Pathfinder suggests the optimal route to 80 drivers, not all of them will follow it), therefore 2) we need to figure out how many drivers to suggest this route to in order to guarantee the right number of cars are around at the right time.
This means we need to consider precise distances to minimize dead mileage and sift out potential routes that are too “costly.”
Pathfinder crunches all these numbers and sends out the optimal number of suggestions. Then when drivers start heading there, we try and find them rides along the way.
Pathfinder Increases Driver Earnings
Pathfinder’s pilot launch showed that drivers who followed its suggestions made on average 20% more than they used to in the same amount of time.
We also noticed that our algorithm worked wonders for drivers new to the service. In the old days, it could take weeks for a new driver to learn the ins and outs of the city and figure out how to maximize their chances of finding their next passenger fast. How else could they know exactly where and when demand spikes? Now Pathfinder gets new drivers feeling confident and secure in just a day or two.
Pathfinder is also the first step to helping drivers plan their entire work day in advance: just set the time you want to work, and the system books it with a step-by-step driving schedule, down to the final paid ride on the way home. In other words, driver earnings are getting more predictable and the stress of unpredictability is falling by the wayside. What more could a driver want than their own personal assistant planning how to make the most money?
But what’s even more exciting is that Pathfinder is so much bigger than just a tool to maximize driver earnings. It’s destined for center stage in the future of self-driving cars.
Today, the race to build the best self-driving car is on, with Alphabet, Uber, Yandex and even car manufacturers all vying for first place. But designing an autonomous vehicle capable of “seeing” its surroundings and driving safely is just half of the equation.
The other half is teaching it how to make navigation decisions: exactly where to go and when to pick up the next passenger and drop them off as quickly as possible. When the roads are full of self-driving cars, the market leader will be whoever knows how to best manage where they’re going.
